feat(tts): add Kokoro-82M TTS model implementation (#127)

[m96-chan]

Summary

• Add pygpukit.tts module with Kokoro-82M StyleTTS2-based TTS model
• Implement native LSTM CUDA kernel (Driver API compatible, no cudaMemcpy)
• Add bidirectional LSTM support with forward/backward concatenation
• Add SafeTensors/PTH weight loading with HuggingFace Hub support
• Add G2P (grapheme-to-phoneme) tokenizer with misaki integration
• Add WAV export/import utilities

LSTM Kernel Implementation

Features

• Kernel-based memory copies - No cudaMemcpy, Driver API compatible
• Stream support - Uses internal::get_capture_stream() for CUDA Graph compatibility
• Bidirectional support - Forward + backward passes with concatenation kernel

Kernels Added

• lstm_cell_f32_kernel - Single LSTM cell computation (gates, cell state, hidden)
• copy_f32_kernel - Simple DtoD copy (replaces cudaMemcpy)
• lstm_copy_to_output_f32_kernel - Strided copy for sequence output
• lstm_concat_bidirectional_f32_kernel - Concatenation for bidirectional output

Python API

import pygpukit as pk

# Unidirectional LSTM
output, h_n, c_n = pk.lstm_forward(x, W_ih, W_hh, b_ih, b_hh)

# Bidirectional LSTM
output, h_n, c_n = pk.lstm_bidirectional(
    x,
    W_ih_fwd, W_hh_fwd, b_ih_fwd, b_hh_fwd,
    W_ih_bwd, W_hh_bwd, b_ih_bwd, b_hh_bwd,
)

Performance

RTX 5090, batch=8, seq_len=100, input=768, hidden=512:

• 39.96 ms/forward
• ~20,000 tokens/sec

TTS Architecture

• PLBERT: 12-layer BERT encoder (768 hidden, 12 heads)
• StyleEncoder: 128-dim style conditioning
• Decoder: 3-layer conv decoder (512 hidden)
• ISTFTNet: 60x upsample vocoder (24kHz output)

Files Added/Modified

• native/ops/nn/recurrent/lstm_kernels.cuh - LSTM CUDA kernels
• native/ops/nn/recurrent/lstm.inl - LSTM dispatch
• native/bindings/nn/recurrent.cpp - Python bindings
• src/pygpukit/ops/nn.py - lstm_forward, lstm_bidirectional
• src/pygpukit/tts/kokoro/ - Kokoro model implementation
• examples/tts.py - LSTM tests and benchmarks

Test plan

• Unidirectional LSTM test passes
• Bidirectional LSTM test passes
• Output shape verification
• Non-zero output validation
• Ruff lint passes
• Mypy type check passes

Closes #127

🤖 Generated with Claude Code