Fix LocalBackend fork_checkpoint to overwrite initial LoRA for vLLM

[arcticfly]

Problem

LocalBackend._experimental_fork_checkpoint has two issues that prevent forked LoRA checkpoints from being used correctly:

1. vLLM loads the wrong checkpoint

model.register(backend) creates an empty LoRA at checkpoints/0000. The fork then copies the real weights to checkpoints/{source_step} (e.g. 0686). But start_openai_server calls get_last_checkpoint_dir() which finds the forked checkpoint — however the vLLM subprocess is configured with @0 alias pointing to 0000.

2. Trainer uses the wrong weights

UnslothService._state is a cached_property that may be initialized before the fork runs. Even when it re-initializes after fork, create_unsloth_train_context calls FastLanguageModel.from_pretrained(model_name=checkpoint_dir) which sets up the LoRA architecture but may not load the trained weights correctly across precision boundaries (e.g. checkpoint trained in 4-bit, loaded in 16-bit).

3. Mixed bf16/fp16 dtype crash on H200

On H200 GPUs, base model activations run in bf16 while LoRA adapter weights are fp16. Unsloth's fused matmul_lora and fast_linear_forward call addmm_/addmv_ which crash on mixed dtypes: RuntimeError: self and mat2 must have the same dtype, but got Half and BFloat16.

Fix

Checkpoint loading (backend.py)

1. Overwrite checkpoints/0000 with the forked weights so vLLM loads the correct adapter on startup.
2. Invalidate the _state cache so the trainer re-initializes with the forked checkpoint path.
3. Set _forked_checkpoint_dir on the service and call load_adapter_from_checkpoint on the first training call to explicitly load the adapter weights via set_peft_model_state_dict.

Dtype patch (dtype_patch.py)

Patches matmul_lora and fast_linear_forward to cast tensors to a common dtype (preferring bf16) before fused ops. Applied automatically when _state is first accessed.

Verification

Without fix: val/qa_failed = 40-80% at step 0 (base model behavior) + dtype crashes
With fix: val/qa_failed = 7% matching W&B Inference baseline (8%), no crashes

pr-test-001 (no local patches, ART PR only):
  val/qa_failed:              7.0%
  val/formatting_failed:     33.0%
  val/filler_words_failed:    4.0%
  val/self_correction_failed: 8.0%
  val/voice_commands_failed:  3.0%
  val/faithfulness_failed:    1.0%
  val/lost_meaning_failed:    4.0%
  val/organization_failed:    1.0%
  val/emoji_failed:           2.0%
  val/reward:                 0.681

Closes #651