[Refactor] Tiler: read the multi-valued aliases set instead of the legacy _alias

[lee2716]

Summary

VariableBuffer already exposes a modern, multi-valued alias set
(self.aliases: Set[str], Deeploy/DeeployTypes.py), but the tiling
machinery (TilerExtension, MemoryScheduler) still reads the legacy
single-valued _alias attribute. As a result, templates that need
alias-based pointer passthrough during tiling have to set the legacy
_alias as a workaround, in addition to populating aliases.

Where the legacy attribute is read

• Deeploy/TilingExtension/TilerExtension.py (~L299, L336, L339)
• Deeploy/TilingExtension/MemoryScheduler.py (~L304, L372, L658-660)

The workaround it forces

• Deeploy/Targets/Generic/Templates/ReshapeTemplate.py — sets
  bufferOut._alias = bufferIn.name on top of bufferOut.aliases.add(...).
• Deeploy/Targets/PULPOpen/Templates/ReshapeTemplate.py — same hack in a
  subclass.

Proposed fix

Make TilerExtension and MemoryScheduler consume buffer.aliases (the set)
instead of buffer._alias. This:

1. supports multiple aliases per buffer (the single-valued _alias cannot);
2. lets us drop the _alias workaround from the Reshape templates.

Scope / notes

• Touches tiling core (DeeployTypes.py, TilerExtension.py,
  MemoryScheduler.py) and therefore affects all platforms, so it belongs in
  its own PR.
• Discussed in PR Add FP32 operators and tiling support for MicroLlama on Snitch  #153 (FP32 MicroLlama on Snitch), where the workaround was
  kept intentionally following the ReshapeTemplate subclass pattern.

[lee2716]

Duplicate of #130 — same leftover _alias / aliases-coverage issue in the tiler.

[lee2716]

Some additional context, and a note that reading aliases in place of _alias is not a direct substitution — the two attributes have different semantics.

Background: the aliasing mechanisms over time

Field	Introduced	Semantics	Where it is used
_alias (single str)	open-source release (predates #51)	directed and chain-resolvable — dealiasBuffer walks derived → … → root until a buffer has no _alias	TilerExtension, MemoryScheduler, dealiasBuffer
alias_of (List[str])	#91 (fixed #30)	directed — "I am an alias of these ancestors", with has_live_ancestors()	the dealloc pass (MemoryAllocation.py)
aliases (Set[str])	#125	symmetric — doubly-linked; has_live_aliases() does a BFS in all directions	liveness (has_live_aliases)

#91 and #125 both changed the buffer classes (DeeployTypes.py) and the dealloc/liveness pass (MemoryAllocation.py); neither modified TilerExtension or MemoryScheduler. As a result the tiler still reads _alias, which is what this issue tracks. (For reference, #30 — "the _alias method used in Reshape kernel fails" — was closed as fixed by #91; it may be worth confirming whether that was purely the dealloc-side failure or also touched the tiler path.)

Why reading aliases in place of _alias is not a direct substitution

The tiler relies on two properties that the symmetric aliases set does not provide:

1. Direction / root. dealiasBuffer needs to find which buffer holds the real allocation (the root); MemoryScheduler distinguishes the aliaser from the aliasee (for example, an in-place output is treated as costless when its input is in the same pattern, and alias buffers are skipped in the worst-case size computation). A symmetric set cannot tell the root from the derived buffer.
2. Acyclicity. dealiasBuffer walks until it reaches a buffer with no _alias. On the symmetric doubly-linked aliases, in ↔ out forms a 2-cycle, so a chain walk over it would trigger the existing "Circular aliasing detected!" assertion.

aliases is symmetric because liveness only needs to know whether any co-allocated buffer is still live, for which an undirected BFS is sufficient. That is also why it does not carry the information the tiler needs.

A fix that covers both mechanisms

The tiler's directed-to-root need and liveness's symmetric-reachability need are two traversals of the same directed graph. Aliasing can be modeled as a directed graph where each buffer records the parent(s) it is derived from (the direction that alias_of originally carried):

• Liveness — ignore direction and BFS both ways. Symmetric reachability over a directed graph is unchanged, so has_live_aliases keeps its current behavior.
• Tiler / dealiasBuffer — follow the directed edges toward the root, as a replacement for _alias, with no cycles.

_alias can then be removed, the _alias assignment in the PULPOpen ReshapeTemplate subclass can go away (it can reuse the Generic template, which already populates only aliases), and a single field serves both subsystems.

Scope

• _alias is written in one production location: Targets/PULPOpen/Templates/ReshapeTemplate.py.
• _alias is read in: DeeployTypes.py::dealiasBuffer, TilerExtension.py (3 sites), MemoryScheduler.py (3 sites), and testSchedulingExtension.py.

Open design questions

1. Direction convention — store child→parent ("parents I'm derived from", matching dealiasBuffer root-finding) or parent→child?
2. One field or two — keep aliases symmetric and add a direction marker, or replace it with a directed field plus an undirected-traversal helper for liveness.
3. Fan-out — _alias is single-valued while aliases is a set; does the tiler ever need a buffer to alias more than one root, and what is the intended behavior if so?
4. Chain depth — does the tiler see multi-hop chains (reshape-of-reshape) in practice, or only depth-1? The replacement should preserve chain resolution in either case.
5. Tests — the reshape tiled tests on Snitch / Siracusa / GAP9 exercise this path: the _alias reads gate aliased-buffer handling, so removing _alias without migrating them changes tiler behavior for reshape and should be validated against those tests. testSchedulingExtension.py also reads _alias and would need migrating.