Support LoRA lifecycle operations for MXFP8 and NVFP4 inference

[Davids048]

Summary

MXFP8 and NVFP4 inference quantize transformer weights after merging the constructor-provided LoRA adapter. LoRA lifecycle operations must update those quantized weights before the next inference request.

PR #1796 rejects lifecycle operations that cannot maintain this invariant. This issue tracks implementing those operations.

Review context: #1796 (comment)

Behavior implemented by PR #1796

Public operation	MXFP8	Default NVFP4	NVFP4 with retained BF16 weights
unmerge_lora_weights()	Explicitly rejected	Explicitly rejected	Explicitly rejected
set_lora_adapter()	Allowed and requantized	Explicitly rejected	Allowed and requantized

Required work

• Support MXFP8 LoRA unmerge by quantizing the restored BF16 weights before inference.
• Support NVFP4 LoRA unmerge by preserving or restoring the BF16 weights and then quantizing the restored weights.
• Support default NVFP4 adapter changes by preserving or restoring the BF16 weights, merging the requested adapter, and then quantizing the merged weights.
• Keep quantized weights synchronized after every merge, unmerge, adapter change, strength change, and accumulated adapter update.

Tests

• Verify MXFP8 unmerge and remerge behavior.
• Verify NVFP4 unmerge with both removed and retained BF16 weights.
• Verify adapter replacement for MXFP8 and NVFP4.
• Compare lifecycle outputs against the equivalent BF16 LoRA execution.