fix: remove partial disk file on aborted uploads

[raashish1601]

Fixes partial disk file cleanup when abort happens after a file write started; also stabilizes fixture-size assertions and temp path checks in tests.

[MohammedAlkindi]

Ran this on Windows 11 (Node 24, default core.autocrlf=true clone — i.e., what a Windows contributor actually gets), both this branch and current main, because CI here has no Windows lane. Two findings, one of them probably not visible from a Linux checkout.

1. The product half of this PR appears superseded on main. 9c801c7 ("fix: clean up in-progress disk writes on abort") landed after this branch was cut and implements the same mechanism this PR adds — a pending-files list, cleanup of in-progress writes on abort, an eager file.path, and a file.stream.destroyed guard. The two edits collide in the same regions of lib/make-middleware.js and storage/disk.js, which is why this PR now shows as conflicting. I compared the diffs side by side rather than inferring from titles.

2. The test half is independently valuable, and the numbers show why. On a default Windows clone:

• current main (4203ccd): 71 passing, 12 failing — every failure is the same shape, e.g. should process parser/form-data POST request: 1803 !== 1778
• this branch (fc39d3b): 73 passing, 0 failing

The mechanism: the .dat fixtures have no .gitattributes entry, so autocrlf smudges them to CRLF at checkout (git ls-files --eol test/files/small0.dat → i/lf w/crlf, 1778 → 1803 bytes on disk), and main's tests assert hardcoded byte sizes (assert.strictEqual(req.file.size, 1778)). This PR replaces the hardcoded sizes with sizes measured from the on-disk fixture, which makes the suite pass on a Windows clone as a side effect — an unadvertised benefit that seems worth keeping regardless of what happens to the storage change.

Disclosure: I have an open PR (#1450) that fixes the same 12 failures at the git level by adding .gitattributes for the fixtures. The two approaches are complementary rather than competing — measured sizes keep the tests robust even where a clone's smudge state is unexpected — but the maintainers should know I'm not neutral on that corner.

Given both points, the highest-value path for this PR might be: rebase over 9c801c7, drop the now-duplicated storage logic, and keep the measured-size test changes (plus the new abort assertions in test/error-handling.js, which ran green here).