feat(train): PEFT (LoRA / LoHa) adapter support for finetune CLI

[medyas]

Closes #621.

Adds optional PEFT (parameter-efficient finetuning) support via the peft library. Defaults preserve current behavior (--peft_method none), so existing full-finetune users are not affected.

What

CLI (src/f5_tts/train/finetune_cli.py) — 4 new flags:

--peft_method [none|lora|loha]     # default: none
--peft_rank N                       # default: 8
--peft_alpha N                      # default: 8
--peft_target_modules \"a,b,c\"     # default: DiT/UNetT attn+FFN linears

Default targets: to_q,to_k,to_v,to_out.0,ff.ff.0.0,ff.ff.2. Excludes (NaN-safe): AdaLN-Zero modulation (attn_norm,ff_norm), final zero-init linears (norm_out,proj_out), and conditioning paths (time_embed,text_embed,input_embed,long_skip_connection).

Trainer (src/f5_tts/model/trainer.py)

• New kwarg peft_config: PeftConfig | None. When set: self.model = get_peft_model(self.model, peft_config) before accelerator.prepare().
• EMA disabled under PEFT (tracking EMA across base + adapter bloats state and degrades adapter learning). All EMA accesses guarded by self.ema_model is not None.
• Optimizer changed from model.parameters() to self.model.parameters() so it sees the wrapped, frozen-base view (this was also a latent bug pre-PEFT — same fix).
• save_checkpoint additionally writes adapter-only safetensors via save_pretrained() to checkpoint_path/adapter_<step>/ for portable, small (~MB) artifacts.
• load_checkpoint tolerates missing ema_model_state_dict and uses strict=False when PEFT is enabled (base key prefix differs after wrap).

CLI pre-load path

When --peft_method != none AND --finetune, the pretrain checkpoint is pre-loaded into the bare CFM before adapter wrap (strict=False). This avoids the Trainer's resume path trying to load un-prefixed pretrained .safetensors into a wrapped (base_model.model.…) model.

Dependency

New optional extra: pip install -e \".[peft]\".

Why this design

• Defaults preserved: every existing finetune command behaves identically. Only users who pass --peft_method lora|loha see new code paths.
• EMA skip over EMA-wrap: EMA on adapters is poorly defined and the marginal benefit doesn't justify the state-management surface area in a first cut. Easy to add as a follow-up.
• Adapter-only artifact: keeps Hub-friendly small files alongside the existing full .pt. Users can ship just the adapter (~MB) once trained.

Validation

• Ruff lint + format clean (matching .pre-commit-config.yaml v0.11.2).
• AST parse of both modified files.
• CPU smoke tests (LoRA + LoHa forward, trainable-param count band, LoHa zero-init identity contract, adapter save/reload roundtrip): 5/5 pass.

GPU smoke (full pipeline at scale) is queued as a follow-up. Happy to gate the merge on it landing if reviewers prefer.

Recommended target_modules / rank for F5-TTS DiT

• LoRA: r=16-32, alpha=r.
• LoHa: r=4-8, alpha=r (LyCORIS rule r <= sqrt(hidden_dim) — F5TTS_Base has dim=1024 so r=8 is a sane sweet spot; effective rank ≈ r² via the Hadamard product with ~½ the LoRA params).

Both honor the same target/exclude conventions to avoid touching AdaLN-Zero modulation or final zero-init linears (which break F5-TTS's init contract and produce NaN within the first updates).