Fix packed sample count derivation

[taking-lying-flat]

Summary

• derive the packed sample count from cu_seqlens_q instead of the nonexistent PackedSeqParams.num_samples field
• fix the fallback Qwen3-Next and Qwen3.5 GDN packed paths
• fix the Qwen3.8 PLE THD unpack path and keep max_seqlen_q compatible with Megatron-Core 0.16.x

This is a minimal follow-up to #175. It does not change packed offsets or add inter-sequence padding behavior.

Checks

• pre-commit run --files src/mcore_bridge/model/gpts/qwen3_next.py src/mcore_bridge/model/mm_gpts/qwen3_5.py src/mcore_bridge/model/modules/ple.py

[tastelikefeet]

Thanks. Fixed by #186