cuda: add paged ragged KV research runtime

[ZacharyZcR]

Summary

• replace grow-copy ragged CUDA KV buffers with fixed 64-token physical pages
• extend ragged attention and append kernels to consume page tables across page boundaries
• add per-device CUDA group telemetry and accurate asynchronous GPU timing
• expose opt-in placement/MTP research controls used by the held-out experiment matrix
• strengthen cross-page and length-skew ragged-attention coverage

Measured behavior

The 64-token page size was the practical allocation/fragmentation Pareto point. The prototype remains numerically exact across page boundaries and reserves 76.4% less KV memory than fixed-slot allocation on the tested length-skew trace. This PR does not claim a decode throughput gain.

Validation

• make -C c check: native suite passed; 292 Python tests passed, 18 skipped
• CUDA compilation is delegated to CI because this workstation has a CUDA GPU but no local nvcc toolkit

Related: #537, #768