bluejay-parser: performance optimizations

[swalkinshaw]

~16% faster schema parsing, ~12% faster executable parsing

Key optimizations:

• Rewrite block string parser: direct string processing instead of sub-lexer + Vec<Vec> (~10%)
• Compact Span: u32 start+len (8 bytes) instead of Range (16 bytes), add Copy (~3%)
• Field: consume-then-check for alias instead of peek(1) (~2%)
• Optimize next_if_* methods: peek+consume in single buffer operation (~1%)
• Lazy depth_limiter.bump(): only bump when optional elements exist
• Add Copy to DepthLimiter + preallocate Vec capacity in DefinitionDocument

Also adds benchmarks for ExecutableDocument parsing with a large fixture, and updates downstream crates to use Copy semantics on Span (clone → deref).

Note: this is a manually curated (with the help of Claude) and cleaned up version of an /autoresearch run

[adampetro]

Should we just change the return type of TryFromTokens::try_from_tokens to return the transposed type and have this be the implementation? I think we call transpose almost every time we call this

[swalkinshaw]

one caller in directives.rs uses Option<Result<...>> 😓 so we could do that but we'd have to update the usage globally. I don't think this is so bad because it makes it obvious that we're manually avoiding calling depth_limiter.bump() when the branch isn't taken?

[adampetro]

Could we change the implementation of TryFromTokens to do the bump internally?

[swalkinshaw]

It becomes a bit messy since there's a few cases which don't bump so it would simplify most of the callers but make a few more complex. Just changing the return type from Option -> Result we can at least get rid of the manual transpose()

[swalkinshaw]

Refactored this a bit

[adampetro]

Why u32 instead of usize?

[swalkinshaw]

Makes the struct smaller so all the .clone() calls become Copy and avoids heap allocations