docs: clarify --ref_audio needs an absolute path

[moduvoice]

Problem

Following the README's CLI example verbatim, but passing a relative path for --ref_audio (the same convention the repo's own bundled infer/examples/basic/basic.toml uses: ref_audio = "infer/examples/basic/basic_ref_en.wav"), fails:

FileNotFoundError: [Errno 2] No such file or directory: '.../site-packages/f5_tts/venv/lib/python3.10/site-packages/f5_tts/infer/examples/basic/basic_ref_en.wav'

The path gets doubled — preprocess_ref_audio_text() resolves the given path relative to the installed f5_tts package directory, not the caller's working directory. An absolute path works immediately with no other change:

f5-tts_infer-cli --model F5TTS_v1_Base \
  --ref_audio "/absolute/path/to/basic_ref_en.wav" \
  --ref_text "Some call me nature, others call me mother nature." \
  --gen_text "I do not really care what you call me. I have been a silent spectator, watching species evolve, empires rise and fall."

Fix

• README — one line noting that --ref_audio (and other file-path CLI args) should be given as an absolute path, since relative paths resolve against the installed package directory rather than the current working directory.

No code/behavior change — this is a documentation-only clarification.

Verification (Tesla T4 16GB)

Reproduced twice with the fix (absolute path) applied, using the maintainer's own bundled example (infer/examples/basic/basic.toml values):

Metric	Run 1	Run 2
Generation time	4.43s	4.50s
Wall time (incl. model+vocoder load)	16.26s	16.32s
Output duration	11.24s	11.24s (deterministic)

Peak VRAM: 1.625 GB (comfortably fits Tesla T4's 16GB, ~90% headroom).

Environment: Tesla T4 16GB · driver 550.163.01 (CUDA 12.4) · Python 3.10.12 · torch 2.8.0+cu128 (the README-pinned cu128 wheel works fine on this driver — no CUDA-version mismatch here).

Also confirmed the model's own dtype auto-selection (load_checkpoint() picking torch.float16 when torch.cuda.get_device_properties(device).major >= 7) already does the right thing on Turing/T4 — no change needed there, this PR is scoped to the path-resolution documentation gap only.

Raw metrics JSON

{
  "run1": {"gen_time_s": 4.43, "wall_time_s": 16.263, "output_duration_s": 11.24, "sample_rate": 24000},
  "run2": {"gen_time_s": 4.50, "wall_time_s": 16.316, "output_duration_s": 11.24, "sample_rate": 24000},
  "peak_vram_mib": 1625,
  "gpu": "Tesla T4", "driver": "550.163.01", "torch": "2.8.0+cu128"
}