Add type-aware extrema and integer ceil-division APIs [wip]

[Phil-amd]

Summary

• Add stable fx.max(), fx.min(), and fx.ceildiv() exports with per-type lowering (float → maximumf/minimumf, signed int → maxsi/minsi/ceildivsi, unsigned int → maxui/minui/ceildivui) and compile-time folding where host semantics match.
• Migrate audited kernel call sites (softmax, top-k gating, RMSNorm, LayerNorm, pa_metadata, shared cdiv helper) to the new typed APIs; mechanical extrema migrations produce byte-identical normalized IR and final ISA.
• Add unit tests, an AST-based CI guard (scripts/check_typed_arithmetic_usage.py), and update authoring docs/skills to steer away from deprecated method-form extrema and raw MLIR builders.

Closes #994. Child of #934.

Test plan

• pytest tests/unit/test_typed_arith_ops.py tests/unit/test_typed_arithmetic_guard.py
• Focused kernel tests: softmax, topk gating, rmsnorm, layernorm (gfx942 + gfx950)
• Full CI on gfx942 (308): 1,755 passed (excluding known-flaky test_kernel_known_block_size.py and multi-GPU tests)
• Full CI on gfx950 (355): passed with same exclusions
• MoE sorting kernel: 61/61 passed in isolated run
• IR/ISA equivalence verified for migrated reduction kernels