[Operator Mechanism] Support FP32 output for FP16/BF16 paddle.bmm on CUDA (#79552)

* [Python API] Move bmm to handwritten implementation

* [Operator Mechanism] Support bmm out_dtype for BF16 CUDA

* [Operator Mechanism] Support bmm out_dtype for FP16 CUDA

* [CodeStyle] Fix bmm out_dtype pre-commit issues

* [API Compatibility] Preserve paddle.tensor.linalg.bmm

* [Test] Improve bmm Python API coverage

* [CodeStyle] Centralize bmm validation in InferMeta

* [Python API] Restore bmm legacy static support

* [Python API] Validate bmm out_dtype device support

* [Python API] Reject bmm out_dtype on ROCm

* [API Compatibility] Align bmm positional signature

* [API Compatibility] Preserve bmm positional name