perf(sax): intern attribute names in the engine hot path

[evoactivity]

What

Interns attribute names in the SAX engine hot path via a small direct-mapped cache (512 slots, cheap length/first/last-char fingerprint, verify loop for collisions).

Why

Attribute names repeat constantly, and every fresh attribute-name substring handed to the keyed attribute store must be re-internalized by V8 before it can be used as a property key. Returning the same string instance for repeated names means repeat occurrences allocate nothing and the store receives an already-internalized key.

Measured results (Node 22, isolated processes, medians)

workload	before	after
@tuananh/sax-parser repo benchmark (attr-heavy synthetic, single write)	25,234 op/s	27,206 op/s (+7.8%)
tokenize: attr-heavy synthetic (256 B chunks)	15,605 op/s	17,834 op/s (+14.3%)
tokenize: XMLTV EPG (256 B chunks)	20,970 op/s	22,401 op/s (+6.8%)
tokenize: RSS / Maven POM	—	neutral

Scope is deliberate

Interning is applied to attribute names only. Two rejected variants, kept out with a code comment so they don't get re-attempted:

• Interning tag names as well: large net loss on element-dense documents (RSS -22%, POM -24%) — tag names are handler arguments, never property keys, so they pay the cache cost for no internalization benefit.
• Full-content hash instead of the cheap fingerprint: -38% overall; the hash loop cost more than the substring allocation it avoided.

Also experimented with (and rejected) a manual short-scan fast path replacing indexOf for attribute values: it only wins on sub-8-char values (the synthetic benchmark shape) and loses ~5-7% on EPG's ~20-char timestamp values — V8's vectorized indexOf is too good at realistic value lengths.

All 838 tests pass.