Fix SSM kernel bugs and unify Mamba prefill through varlen path

[lmcafee-nvidia]

Summary

• conv_state save-before-read: extract initial conv states BEFORE causal_conv1d_varlen_states + tensor_masked_update overwrites the conv_state buffer, fixing corrupted convolution output for restored requests
• cu_chunk_seqlens OOB: subdivide each sequence into chunks of at most chunk_size tokens instead of passing cu_seqlens directly, fixing out-of-bounds memory access when sequences exceed chunk_size (128) tokens
• zxBCdt padding mismatch: strip padded tokens from zxBCdt before _ssm_prefill, then pad the output back, fixing shape assertion failures when padded and real token counts differ
• per-request conv1d with initial_states: loop over requests calling causal_conv1d_fn per-request with initial_states, because causal_conv1d_fn cannot accept both seq_idx and initial_states simultaneously
• unify prefill path: all Mamba prefill (including chunked) goes through a single varlen SSM kernel call, removing separate chunked-prefill routing and _batch_indices_chunked_prefill / _device_chunked_prefill metadata
• simplify _dynamic_inference: flat decode + prefill structure with new _dynamic_inference_prefill helper that strips CUDA-graph padding from metadata and data tensors
• remove deprecated constructor params: use_mem_eff_path, d_state, headdim, ngroups and their warnings.warn blocks
• add assertion format string in ssd_combined.py for easier debugging
• add batch_allocate_slots() to MambaMetadata for batch slot allocation

Test plan

• verify mamba_mixer.py imports without errors
• run existing SSM kernel unit tests
• run hybrid model inference end-to-end with decode + prefill mix
• verify chunked prefill requests work through the unified varlen path

🤖 Generated with Claude Code