fs: allocate FSReqPromise stat arrays lazily

[MarshallOfSound]

Description of Change

Every promise-based fs operation allocates two AliasedBuffers at request construction — an 18-field stats array and an 8-field statfs array — although only stat-family resolutions ever read the first and only statfs() reads the second. Each one costs an ArrayBuffer, a TypedArray and a strong v8::Global, so a plain fsp.writeFile() pays for two backing stores it can never use. The callback path has no equivalent cost: FSReqCallback resolves stats through a shared per-process array.

Under concurrency this allocation churn is the dominant overhead of the promises path — profiling 64 in-flight 1KB reads shows ~21% of CPU ticks in the malloc family for promises vs ~5% for callbacks issuing identical syscalls.

This PR constructs the arrays lazily in ResolveStat()/ResolveStatFs():

• Non-stat operations (open, read, write, close, unlink, …) now allocate neither array; stat-family ops allocate one; only statfs allocates the statfs array.
• Once created, lifetime is unchanged — the array is still owned by the request, so deferred promise continuations read from request-owned memory exactly as before. The reason the per-request copy exists (continuations, unlike callbacks, don't consume the result synchronously) is preserved.
• MemoryInfo reports the arrays only when they exist.

benchmark/compare.js shows significant improvements on the concurrent configurations — writefile-promises +12.6…20.7%, readfile-promises +19.1…21.5% (unencoded), bench-stat-promise +2.2…3.5% — with sequential large-payload configs neutral (those are threadpool-latency-bound, as expected).

Benchmark results

benchmark/compare.js, 10 samples per configuration, Welch t-test (*** p<0.001, ** p<0.01, * p<0.05):

fs/bench-stat-promise.js statType='fstat' n=200000                          +3.51% ***
fs/bench-stat-promise.js statType='lstat' n=200000                          +2.20% *
fs/bench-stat-promise.js statType='stat' n=200000                           +2.88% ***
fs/readfile-promises.js c=1  len=1024     enc=''       +3.25% ***   enc='utf-8'  +4.21% ***
fs/readfile-promises.js c=1  len=512k     enc=''       +3.61%       enc='utf-8'  +4.38% **
fs/readfile-promises.js c=1  len=8M       enc=''       +8.34% **    enc='utf-8'  -2.35%
fs/readfile-promises.js c=1  len=16M      enc=''       +2.30%       enc='utf-8'  +1.41%
fs/readfile-promises.js c=1  len=32M      enc=''       -2.17%       enc='utf-8'  +0.43%
fs/readfile-promises.js c=10 len=1024     enc=''      +21.02% ***   enc='utf-8' +19.84% ***
fs/readfile-promises.js c=10 len=512k     enc=''       +7.07% *     enc='utf-8'  +0.94%
fs/readfile-promises.js c=10 len=4M       enc=''      +19.14% **    enc='utf-8'  +1.39% *
fs/readfile-promises.js c=10 len=8M       enc=''      +21.51% ***   enc='utf-8'  -0.25%
fs/readfile-promises.js c=10 len=16M      enc=''      +20.99% ***   enc='utf-8'  +0.99% *
fs/readfile-promises.js c=10 len=32M      enc=''       +5.16% **    enc='utf-8'  +0.76%
fs/writefile-promises.js c=1  size=2      asc +7.17%*** buf +6.00%*** utf +5.83%***
fs/writefile-promises.js c=1  size=1024   asc +5.19%*** buf +6.00%*** utf +5.06%***
fs/writefile-promises.js c=1  size=64k    asc +2.65%*** buf +4.00%*** utf +2.12%***
fs/writefile-promises.js c=1  size=1M     asc +0.88%    buf +0.93%*   utf -0.38%
fs/writefile-promises.js c=10 size=2      asc +19.80%*** buf +18.70%*** utf +20.71%***
fs/writefile-promises.js c=10 size=1024   asc +13.51%*** buf +18.77%*** utf +12.58%***
fs/writefile-promises.js c=10 size=64k    asc +4.58%***  buf +2.65%**   utf +11.02%***
fs/writefile-promises.js c=10 size=1M     asc +2.25%*    buf +0.15%     utf +0.06%

Additional ad-hoc microbenchmarks (tmpfs, 64 ops in flight): fsp.writeFile 1KB +53%, fsp.stat +26%, fsp.readFile 1KB +22% — with all callback variants unchanged (±3%), confirming attribution since callbacks never construct FSReqPromise. Scripts: https://gist.github.com/MarshallOfSound/7d22707e3e550a782e96f9da0081e9f1

Checklist

• I have built and tested this change
• test/parallel/test-fs-* pass (342 tests; 3 pre-existing environment failures fail identically on main)
• Verified fsp.stat, fsp.statfs and fsp.readFile resolve correctly through the lazily-created-array paths

[addaleax]

Not a blocker for this PR, but we should support std::optional in MemoryTracker instead of manually handling this case (e.g. add an overload like

node/src/memory_tracker.h

Line 216 in d097de8

const std::pair<T, U>& value,

that transparently covers optional values) -- I don't think it would be a bad idea to add it as a separate commit in this PR, though, if you want to do that

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74107/

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74406/

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74511/