Fix: Align nn.Embedding output with AMP autocast precision

[Xiaoming-AMD]

Summary

This PR fixes a precision mismatch issue where nn.Embedding outputs remain in fp32
even when AMP (autocast) is enabled with bf16 or fp16.
This inconsistency causes unnecessary dtype conversions and increased memory usage.

---

✨ What’s Changed

• Globally monkey-patched nn.Embedding.__init__ to register a forward hook.
• The hook:
  • Checks if AMP (torch.is_autocast_enabled()) is active.
  • If yes, casts the embedding output to the current autocast dtype (e.g., bf16).
• Controlled via environment variable:

 --primus_turbo.enable_embedding_autocast=false  # disables the patch