Make raise_error compose with retriable status retries

[lineoffligbot]

Problem

The raise_error feature and the retriable feature silently conflict. Enabling both makes retry_statuses dead configuration:

HTTP.use(:raise_error)
    .retriable(tries: 3, retry_statuses: [503])
    .get(url)
# expected: 3 attempts, then OutOfRetriesError
# actual:   1 attempt, immediate StatusError

raise_error converts an error response into a StatusError inside each attempt (Client#perform_once wraps the response via feature.wrap_response), so Retriable::Performer catches an exception instead of seeing a response. StatusError is not in RETRIABLE_ERRORS, so retry_request? declines and the error is raised on attempt 1. Measured against a local always-503 server:

configuration	requests made	raised
retriable alone (retry_statuses: [503])	3	OutOfRetriesError with response
plus raise_error	1	StatusError
workaround: StatusError in exceptions:	3	OutOfRetriesError with nil response
raise_error with ignore: covering every retry status	3	OutOfRetriesError with response

The last row composes correctly but requires keeping two lists in sync by hand forever.

Fix

try_request now recovers the response a StatusError carries:

rescue Exception => e
  err = e
  res = e.response if e.is_a?(StatusError)
end

and retry_request? treats a status-matching StatusError as retriable:

elsif err
  retry_exception?(err) || retriable_status_error?(err)

Populating res early repairs the whole downstream pipeline at once: retry_statuses matching, Retry-After delay calculation, the on_retry callback, response flushing on exhaustion, and OutOfRetriesError#response and #cause. A StatusError whose status is not covered still raises immediately on attempt 1, preserving raise_error semantics. This also closes the nil-response hole in the exceptions: workaround (third row above).

The underlying principle: raise-on-status should act only after retry-on-status is exhausted, but raise_error runs per attempt inside the retry loop, so the performer has to translate the exception back into its response.

Notes on observable changes

• on_retry and should_retry callbacks now receive the error response alongside the StatusError (previously the response argument was nil in that case).
• OutOfRetriesError#response is populated when retries were driven by a StatusError, and #cause is the StatusError.

Ten new tests cover the matrix rows, the preserved no-retry paths, StatusError subclasses, delay calculator and callback plumbing, and response flushing, all written to fail before the fix. RBS signature included. Mutant kills 100% of mutations on the diff.

[lineoffligbot]

A note on the failing steep job: it is environment drift, not this patch. The Gemfile does not lock rbs (gem "rbs", ">= 4"), so CI resolves it fresh on every run. The last green run on main (July 14) resolved rbs 4.0.3; runs today resolve rbs 4.1.3, which introduces 8 new warnings in files this PR does not touch, and bundle exec rake steep fails identically on an untouched checkout of main. Happy to send a small separate PR fixing those warnings so CI goes green again.