module: avoid allocating a cache key string for every require()

[MarshallOfSound]

Description of Change

Every require() call — including fully cached ones — allocates a fresh `${parent.path}\x00${request}` string to key the relative-resolve fast path in Module._load. On cache hits that allocation and its GC churn are the largest per-call cost.

This PR splits the cache into two levels so hits do two allocation-free lookups instead: an outer SafeMap keyed by parent.path (a string the module system already retains, with a V8-cached hash) and an inner { __proto__: null } object keyed by the raw request.

• The inner level is a plain object rather than a Map on purpose: property access internalizes the key, which handles dynamically built specifiers (require(base + name)) far better than Map.get's per-call hash-and-compare — and it preserves the string coercion the old template literal performed.
• Invalidation is unchanged: stale hits delete the inner entry, failed loads clean up in the existing finally, and delete require.cache[...] + re-require behaves as before.
• Memory is neutral (12.97 → 12.91 MB heap after loading a 2,000-directory tree) since the concatenated keys are no longer retained.

Observable edge: parent.path is now an identity Map key rather than being string-coerced, and is read twice on the first load into a directory — only reachable if module.path is replaced with a non-string or an accessor, neither of which is documented usage.

benchmark/compare.js shows significant wins on every relevant module-loader.js config — cached +7.8…10.7%, cold +5.6…8.0% — with clean nulls on the circular/deep benchmarks where this path isn't hot.

Full benchmark results

10 samples per configuration, Welch t-test (*** p<0.001, ** p<0.01, * p<0.05):

module/module-loader.js cache='true'  n=1000 files=500 dir='abs' name=''            +10.72% ***
module/module-loader.js cache='true'  n=1000 files=500 dir='abs' name='/'            +8.08% ***
module/module-loader.js cache='true'  n=1000 files=500 dir='abs' name='/index.js'    +7.79% ***
module/module-loader.js cache='true'  n=1000 files=500 dir='rel' name=''            +10.30% ***
module/module-loader.js cache='true'  n=1000 files=500 dir='rel' name='/'           +10.20% ***
module/module-loader.js cache='true'  n=1000 files=500 dir='rel' name='/index.js'    +9.24% ***
module/module-loader.js cache='false' n=1000 files=500 dir='abs' name=''             +6.20% ***
module/module-loader.js cache='false' n=1000 files=500 dir='abs' name='/'            +6.18% ***
module/module-loader.js cache='false' n=1000 files=500 dir='abs' name='/index.js'    +7.89% **
module/module-loader.js cache='false' n=1000 files=500 dir='rel' name=''             +8.01% ***
module/module-loader.js cache='false' n=1000 files=500 dir='rel' name='/'            +5.61% *
module/module-loader.js cache='false' n=1000 files=500 dir='rel' name='/index.js'    +6.79% ***
module/module-require.js n=10000 type='.js'                                          +3.06% **
module/module-require.js n=10000 type='.json'                                        +2.67% ***
module/module-require.js n=10000 type='dir'                                          +0.02%
module/module-loader-circular.js n=1000                                              +0.22%
module/module-loader-deep.js cache='true' files=1000 (both exts)                  ±1.9% (n.s.)

Ad-hoc workload-shape microbenchmarks land between −9% (300 dynamic-specifier children) and −58% (cached-require hot loop) load-phase time, neutral on cold loads — scripts and fixture generators: https://gist.github.com/MarshallOfSound/ad35b07ad5ebf9c5fec8e4daa4ec04b4

Checklist

• I have built and tested this change
• test/parallel/test-module-*, test-require-*, test-cjs-* pass
• Manually verified cache-hit identity, delete require.cache + re-require, failed-load retry, circular requires, require.resolve with paths

[nodejs-github-bot]

Review requested:

• @nodejs/loaders

[mcollina]

lgtm