feat(mudlark): expose structural mutation counters

[da2ce7]

Summary

Release Mudlark 1.1.0 with monotonic counters for structural splits, evictions and restorations, plus a reader for the current semi-internal population. The API change is additive.

Motivation

Consumers can observe the live graph shape, but cannot distinguish accumulated structural churn from the current state: reconstructing split, eviction and restoration history from node-state deltas is expensive and unreliable, and a counter epoch that starts at construction makes interval deltas composable across operations. The semi-internal population is a current-state diagnostic in its own right and had no reader.

Change

• Add StructuralMutationCounts with public splits, evictions and restorations totals; each saturates at u64::MAX rather than wrapping, so it never decreases during the graph's lifetime. A split is counted per child created, so a bisection adds two and a legacy promotion that recreates one missing child adds one restoration.
• Increment the matching counter at every site that changes the node count: bootstrap_split, catalytic_split, evict_tip and handle_legacy_promotes. The audit of node_count mutation sites finds exactly these four.
• Add GvGraph::structural_mutation_counts() and GvGraph::semi_internal_count(), the latter computed by scanning the live nodes because the transitions that create and remove a semi-internal node are spread across splitting, eviction and restoration.
• Cover bootstrap splits, catalytic splits, eviction, legacy restoration and terminal-count behaviour with counter assertions, including a direct restoration test.
• Document why a depth buffer wide enough to overflow the exponent saturates the headroom figure at a ceiling no budget can clear.
• Release as Mudlark 1.1.0: the package version, the lock entry, and the dated changelog section with its comparison link.

Verification

At the head commit:

• cargo +stable clippy --keep-going --workspace --all-targets --all-features -- -D warnings — exit 0.
• cargo +stable test -p torrust-mudlark --all-targets --all-features — exit 0 (668 unit tests, the integration targets, 116 doctests).
• cargo +1.90.0 check -p torrust-mudlark --all-features — exit 0.
• cargo test -p torrust-mudlark --doc --all-features — exit 0.
• RUSTDOCFLAGS="-D warnings" cargo doc -p torrust-mudlark --no-deps — exit 0.
• cargo metadata --format-version 1 --locked — exit 0 and resolves torrust-mudlark 1.1.0 from the committed lockfile.
• Fails-before witness: the base overlaid with only the three test files fails to compile on the absent type, reader and field, linking an isolated base rlib; the head passes the same tests.
• Mutation witness: removing the eviction increment makes terminal_count_decreases_on_eviction fail with left: 0, right: 1; restoring the file reproduces its blob, mode and mtime exactly.

The public-surface diff is exactly the new struct, its three fields, the two readers and the crate-root re-export; nothing public changed or moved. The lock moves only the torrust-mudlark entry. No dependency change.

Review notes

Nightly workspace clippy on the newest nightly reports two double_must_use findings in index-cli-common, both present at the base and outside this change; stable workspace clippy is green and is the lint this workspace certifies on.