perf: optimize OneTrans BF16 training on TF2.15

[weloMThreads]

Summary

• Add OneTrans BF16 fast paths and graph rewrites for attention, RMSNorm, softmax, dropout, gather/cast, and SGD update patterns.
• Restore the minimal TF2.15 MUSA device/runtime registration needed by these kernels.
• Keep the branch scoped to the performance path; no training script changes are included.

Verification

• Built build/libmusa_plugin.so on CI from perf/onetrans-bf16-effective-clean.
• Verified TensorFlow 2.15 plugin load exposes /device:MUSA:0 with MUSA_VISIBLE_DEVICES=6.
• Ran OneTrans BF16 perf test on card 6:
  • Average ms/full step: 41.852
  • Measured steps: 40
  • Last measured train_loss: 0.8828