Remove non-blockwise 8-bit optimizer and legacy quantization CUDA/HIP kernels

[matthewdouglas]

Summary

Follows up on #1871 (which removed the Python-side APIs) by deleting the corresponding C++, CUDA, and HIP implementation code. No Python callers remain for any of this.

What was removed

CUDA/HIP kernels (csrc/kernels.cu, csrc/kernels.hip):

• kPreconditionOptimizerStatic8bit1State
• kOptimizerStatic8bit1State
• kPreconditionOptimizerStatic8bit2State
• kOptimizerStatic8bit2State
• kPercentileClipping<float/half, 2048, 4>
• kQuantize / kDequantize (global/non-blockwise variants)
• All associated #define macros and explicit template instantiation blocks

Ops-level launchers (csrc/ops.cu, csrc/ops.hip):

• optimizerStatic8bit<T, OPTIMIZER>
• percentileClipping<T>()
• quantize() / dequantize()

Headers (csrc/kernels.cuh, csrc/kernels_hip.cuh, csrc/ops.cuh, csrc/ops_hip.cuh):

• All declarations for the above

Python interface (csrc/pythonInterface.cpp):

• MAKE_FUNC8 + 8 instantiations (adam/momentum/rmsprop/lion × fp32/fp16)
• MAKE_CFUNC8 + 8 C-exported wrappers (cadam_static_8bit_grad_32/16, etc.)
• cquantize / cdequantize
• cpercentile_clipping_g32 / cpercentile_clipping_g16

The blockwise variants (optimizerStatic8bitBlockwise, kOptimizerStatic8bit1StateBlockwise, kOptimizerStatic8bit2StateBlockwise) are untouched.

Agent guide updates

Removed stale references to percentile_clipping, block_wise parameter, optimizer_update_8bit (non-blockwise), and the non-blockwise dispatch branch from agents/api_surface.md, agents/architecture_guide.md, and agents/security_guide.md.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

PR Review: #1880 — Remove non-blockwise 8-bit optimizer and legacy quantization CUDA/HIP kernels

Follow-up to #1871: mechanically removes the C++/CUDA/HIP implementations of functions whose Python-side APIs were already deleted. The scope is exactly right — only code with confirmed-zero Python callers is deleted, and the blockwise variants are untouched throughout.

No blocking issues.

What was checked:

• Python caller verification: grep across the entire Python tree confirms zero remaining references to all removed symbols (optimizer_update_8bit, percentile_clipping, cquantize, cdequantize, cadam_static_8bit_*, etc.). The deletions are safe.

• Blockwise path untouched: optimizerStatic8bitBlockwise, kOptimizerStatic8bit1StateBlockwise, kOptimizerStatic8bit2StateBlockwise, and all their header declarations/instantiations remain in place. Verified in both kernels.cu and kernels.hip.

• 32-bit variant not accidentally removed: The remaining kPreconditionOptimizer32bit1State / kPreconditionOptimizer32bit2State references in the diff are the 32-bit ops-path kernels, which are distinct from the deleted 8-bit static ones. No over-deletion.

• CUDA/HIP symmetry: Both backends receive identical removals (quantize/dequantize launchers, optimizerStatic8bit, percentileClipping, template instantiations, header declarations). ✓

• The +1 in csrc/ops.cu: Fixes indentation on MAKE_optimizerStatic8bitBlockwise(half, ADAM); which was over-indented inside the now-removed MAKE_optimizerStatic8bit block. Correct.

• Agent docs: api_surface.md (parameter table and deprecated-symbols table), architecture_guide.md (ops list and dispatch pseudocode), and security_guide.md (trusted hot paths) are all consistently updated. The architecture guide's dispatch pseudocode now accurately shows only the uint8 → blockwise branch.

• Security: Clear (internal author, no new dependencies)

• Downstream impact: None (Python API already removed in #1871; no new breakage)

• Tests: Adequate (no Python callers → no test changes needed; full suite passes)

• CI: All pass (CUDA 11.8–13.x, ROCm 6.2–7.2, CPU x64/aarch64/macOS/Windows, XPU, Lint)

• Serialization: Not affected

• torch.compile: Not affected