docs: add VRAM requirements + note SDPA attention dispatch gap

[moduvoice]

Motivation

While verifying this project's inference pipeline on a real NVIDIA Tesla T4 (16GB, Turing/sm_75 — a
common budget/free-tier GPU, e.g. Google Colab's free tier), I found some hardware-specific
behavior and a documentation gap that aren't currently covered. This PR adds that documentation
only — no code or behavior changes.

Changes

• Add a "Hardware Requirements" section (previously absent): checkpoint download size (model.pt + model-svc.pt = 5.61GB; separate Preprocess bundle 6.92GB), measured peak inference VRAM (1.6–2.7GB), and measured real-time factor (RTF ≈ 0.66) — both SVS and SVC comfortably fit a 16GB T4.
• Note in the Quick Start section that cli.inference has no --lyrics/--text argument — it only accepts pre-built, frame-aligned metadata JSON via --prompt_metadata_path/--target_metadata_path, so synthesizing custom lyrics currently requires running the full preprocess pipeline (vocal separation + ASR + MIDI transcription, ~6.92GB extra checkpoints) first.
• Document a measured attention-backend config/reality mismatch: DiffLlama's top-level config._attn_implementation reports "sdpa", but all 22 transformer layers actually instantiate eager LlamaAttention at runtime (each layer builds its own LlamaConfig(...) in soulxsinger/models/modules/llama.py, which never passes through PreTrainedModel._autoset_attn_implementation(), so it silently falls back to the "eager" class default). Forcing the same weights through real SDPA (self_attn.__class__ swapped to LlamaSdpaAttention, no weight/architecture change) measured 2.31x faster inference and ~40% lower peak VRAM with no measurable output-quality regression. This looks unintentional and may be worth fixing in the attention dispatch code — flagging it here as a candidate issue/PR rather than fixing it in this docs-only change.

Testing

• Measured on a real Tesla T4 16GB (Turing, sm_75), driver 550.163.01, torch 2.2.0+cu121, transformers 4.41.2.
• Checkpoint sizes confirmed via the HF Hub API (model.pt 2,818,092,278 bytes, model-svc.pt 2,793,965,154 bytes; Preprocess bundle 6,917,669,445 bytes across 8 files).
• VRAM/RTF measured running the README's own bash example/infer.sh verbatim (fp16, n_steps=32, cfg=3, control=score): 51.24s of audio synthesized in 33.77s wall time (RTF ≈ 0.659).
• Attention finding verified by (1) runtime inspection — self_attn.__class__.__name__ across all 22 layers is LlamaAttention, not LlamaSdpaAttention, despite config._attn_implementation == "sdpa" — and (2) a controlled before/after benchmark (2 runs each, torch.cuda.synchronize() timing + torch.cuda.max_memory_allocated()): eager 34.05s/34.61s (avg 34.33s, peak VRAM 2.71GB) vs. SDPA-patched 14.81s/14.92s (avg 14.87s, peak VRAM 1.63GB) — 2.31x speedup, ~40% VRAM reduction, output magnitude within normal run-to-run variance.