[Feature][QDP] IQP kernel fusion and persistent kernel / grid-stride optimizations

[400Ping]

Related Issues

Closes #1015

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

Optimize the QDP IQP CUDA path by reducing kernel-launch overhead and global-memory traffic, and by making large single-sample launches scale with a capped grid + grid-stride loop pattern.

How

CUDA kernel changes

• Added grid-stride looping to the single-sample IQP phase kernel
• Added capped-grid launch sizing using MAX_GRID_BLOCKS
• Added a shared-memory fused path for n <= FWT_SHARED_MEM_THRESHOLD
  • phase generation in shared memory
  • full FWT in shared memory
  • normalization on final write-back
• Fused the final global-memory FWT stage with normalization for n > threshold
• Fused batch IQP phase + normalization so batch IQP no longer needs a standalone normalize kernel launch

Tests

Added GPU tests covering:

• FWT/fused IQP vs naive CPU reference for small qubit counts
• batch IQP encode vs repeated single-sample encode
• large-state capped-grid / grid-stride regression (num_qubits = 20)

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes