fix(gralora): use _cast_input_dtype in forward instead of raw .to()

[Chessing234]

Bug

Linear.forward in src/peft/tuners/gralora/layer.py casts the input tensor with x.to(gralora_dtype) in two places. This bypasses the disable_lora_input_dtype_casting context manager provided by BaseTunerLayer, which lets callers suppress dtype casting during specific operations (e.g. AMP, mixed-precision inference).

Root cause

BaseTunerLayer._cast_input_dtype checks self._disable_lora_input_dtype_casting and is a no-op when the context manager is active. A bare .to(dtype) call has no such guard.

Fix

Replace both raw casts with the inherited helper:

# before (line 367)
dropout(x.to(gralora_dtype)).view(...)
# after
dropout(self._cast_input_dtype(x, gralora_dtype)).view(...)

# before (line 382)
gralora_B_general(gralora_A_general(dropout(x.to(gralora_dtype))))
# after
gralora_B_general(gralora_A_general(dropout(self._cast_input_dtype(x, gralora_dtype))))

This is the same pattern already applied to vera (#3172), poly (#3177), fourierft (#3170), and pvera (#3208). gralora was missed in those fixes.