feat: refactored online model to use t5 architecture

[schennam714]

Summary

Replaces the custom vanilla decoder-only transformer in the online model with HuggingFace's T5Stack.

Changes

Architecture (src/musicagent/models/online.py)

Removed:

• RMSNorm, RelativePositionBias, SelfAttentionWithRelPos, DecoderBlock — custom vanilla components

Added:

• T5Stack configured as decoder-only (is_decoder=True, is_encoder_decoder=False)
• Explicit attention_mask parameter in forward() for correct batched generation
• Cumulative attention mask tracking in generate() to handle variable-length padded batches

Preserved:

• Separate melody/chord embeddings with is_melody mask
• Interleaved sequence format [SOS, y₁, x₁, y₂, x₂, ...]
• Public API: forward(), generate(), enable_gradient_checkpointing()

Tests (tests/test_model.py)

• Replaced test_online_causal_mask → test_online_causal_behavior (tests observable behavior, not internal implementation)
• Added test_online_generate_with_variable_length_batch (verifies attention mask fix)

Breaking Changes

• Not checkpoint-compatible with vanilla model weights (fc_out now has bias=False to match T5 convention)

Notes

• Default FFN is ReLU (same as vanilla). Set feed_forward_proj="gated-gelu" in T5Config for T5 1.1 gated variant.
• Cross-attention layers exist in T5Stack but are skipped (no encoder hidden states provided)
• All 55 tests pass