[VirgilMing]

This PR fixes #6498.

Before this PR, read_reparse_point() passes MAX_PATH to the UTF-16-to-UTF-8 conversion, while both callers already allocate MAX_LONG_PATH bytes.
The conversion includes the terminating NUL, so the existing code fails with ERANGE (Result too large) for targets of 260 or more UTF-8 bytes.
The failure affects both readlink() and mingw_lstat() independently of core.longpaths; multibyte targets can reach that byte limit with far fewer than 260 characters.

• This PR changes the conversion's capacity argument to MAX_LONG_PATH, matching the buffers that the callers already allocate.
• This PR also adds one setup test and six regression cases to the existing t2031-checkout-long-paths.sh script.
  • The new cases cover ASCII and multibyte targets at 259, 260 and 261 UTF-8 bytes.
  • Each new case checks the decoded target and requires both the diff output and stderr to be empty.
  • The new tests disable the filesystem cache to exercise mingw_lstat() and set core.longpaths=false to demonstrate that the short link pathname does not require that option.
• The existing buffer allocations and readlink() truncation behavior remain unchanged.

I built the unmodified main code (5cf10d74486d1e59b9ae6cedaedb4e26314e1557) and the patched code on native Windows 11 with MSYS2 UCRT64 GCC 16.2.0.
I used DEVELOPER=1, CFLAGS='-O0 -g' and NO_RUST=1, and disabled optional curl, expat, OpenSSL, gettext, iconv, Perl, Python and Tcl/Tk features for these local validation builds.
Those build settings are not changes in this PR.

I performed the following checks:

• I ran the expanded t2031-checkout-long-paths.sh suite against both binaries. Against the unmodified code, four of this PR's new cases fail with Result too large at 260/261 bytes. Against the patched code, all 14 tests pass with no skips: the seven pre-existing tests, the new setup test and the six new regression cases.
• I reran the existing PowerShell reproducer, which was published before this PR and is not modified or added to the repository by this PR. Its 13 cases complete with both binaries. The unmodified binary produces eight conversion errors on stderr; the patched binary produces no errors or differences. Both script processes exit 0, so I checked stderr explicitly.
• I ran five other existing, unchanged test scripts against the patched code: t0060-path-utils.sh, t2007-checkout-symlink.sh, t2102-update-index-symlinks.sh, t4011-diff-symlink.sh and t7515-status-symlinks.sh. Together, they execute 219 tests successfully and skip 20 according to their prerequisites. This host's shell ln lacks native symlink support; this PR's new regression cases use Windows mklink, and all six execute successfully.