Intern Ident<Canonical> strings via salsa::interned to reduce hashing and cloning overhead

[bpowers]

Context

Profiling full-model compilation of WRLD3 with callgrind and DHAT shows that SipHash operations on Ident<Canonical> strings account for ~14% of all instructions in project_build, and String::clone on these identifiers adds another ~5%.

Ident<Canonical> wraps an owned String today. Every HashMap lookup hashes the full string content, every clone allocates a new String, and every BTreeSet operation does byte-by-byte comparisons.

Proposal

Change Ident<Canonical> to use #[salsa::interned] so that each unique canonical string is stored once and represented by a lightweight integer ID. This converts:

• Hashing: from SipHash over N bytes to hashing a u32/u64 integer
• Cloning: from String::clone (allocation + memcpy) to Copy of a small integer
• Equality: from byte-by-byte string comparison to integer ==
• Ordering: would need a custom Ord impl that dereferences to the underlying string, since integer ordering doesn't match lexicographic ordering

Scope

• Add a #[salsa::interned] newtype (e.g. InternedIdent) with the canonical string as its field
• Migrate Ident<Canonical> usages in the compilation pipeline to use the interned type where a &dyn Db is available
• Ident<Canonical> may still be needed at the edges (parsing, serialization) where there is no salsa database
• Update HashMap<Ident<Canonical>, ...> and BTreeSet<Ident<Canonical>> to use the interned type

Acceptance Criteria

• Ident<Canonical> in the compilation pipeline uses salsa interning
• cargo bench -p simlin-engine --bench compiler -- wrld3 shows measurable improvement in project_build (target: reduce SipHash + String::clone overhead by >50%)
• cargo test -p simlin-engine passes
• DHAT allocation count for full pipeline is reduced (fewer String allocations)
• No behavioral changes in simulation output

Notes

This is a larger refactor. It should be done after the lower-effort canonicalize() cleanup and per-variable unit context caching, which address the most acute regressions from incremental compilation. The interning work is a further optimization on top of those.

Profiling baseline (WRLD3 full pipeline, main branch after incremental compilation):

• SipHash write + finish + c_rounds: ~16.6M instructions (7.8% of total)
• String::clone: ~2.3M instructions
• Total instruction count: 212.4M