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

[Chessing234]

Bug

Linear.forward in src/peft/tuners/pvera/layer.py casts the input tensor with x = x.to(lambda_d.dtype). This bypasses the disable_lora_input_dtype_casting context manager provided by BaseTunerLayer, which is intended to let 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 the raw cast with the inherited helper:

# before
x = x.to(lambda_d.dtype)
# after
x = self._cast_input_dtype(x, lambda_d.dtype)

This is the same pattern already applied to vera (#3172), poly (#3177), and fourierft (#3170); pvera was the only tuner that still used the raw .to() call.

[BenjaminBossan]

@Chessing234 Could you please apply all changes of this type in a single PR instead of opening one for each PEFT method individually?