ptx-schedule: classify tensor-map mutations as schedule sites

[uurl]

Closes #1208.

Summary

Classify tensormap.replace.* instructions as schedule-sensitive sites in ptx-schedule.

Tensor-map replacement operations mutate the tensor-map descriptor through the generic proxy. They must be observable by schedule perturbation independently from the fence.proxy.tensormap::* operation that publishes those mutations.

This adds a dedicated SiteKind::TensorMapMutation rather than classifying these instructions as AsyncProxy or OrderedMemory.

Changes

• Add SiteKind::TensorMapMutation.
• Classify tensormap.replace.* instructions as TensorMapMutation.
• Keep fence.proxy.tensormap::* classified as Fence.
• Add coverage for emitted tensor-map replacement spellings.
• Add a perturbation test verifying that a tensor-map mutation can be delayed before its publish fence.
• Document the new site kind in the ptx-schedule README.

Validation

Validated with:

cargo fmt --all -- --check
cargo test -p ptx-schedule tensormap -- --nocapture
cargo test -p ptx-schedule
cargo clippy -p ptx-schedule --all-targets -- -D warnings
RUSTDOCFLAGS="-D warnings" cargo doc --no-deps -p ptx-schedule
git diff --check

Results:

• 3 tensor-map-specific tests passed.
• 22 ptx-schedule unit tests passed.
• 3 CLI integration tests passed.
• Clippy passed with warnings denied.
• Rustdoc passed with warnings denied.

The original and perturbed PTX were also assembled successfully with ptxas.

--list-sites reports:

tensormap.replace.*       -> TensorMapMutation
fence.proxy.tensormap::*  -> Fence