fix(deps): fix fmtconv's non-x86 integer scaler, pin fmtconv to r31

[StuartCameronCode]

Build-side half of the QTGMC preset-brightening fix on Apple Silicon. Only affects how the deps bundles are built — the app-side pointer bump, tests and docs follow separately once deps-v1.7.0 is published.

Root cause

fmtconv's Scaler::process_plane_int_cpp applied the same sign-conversion constants as process_plane_int_sse2/avx2. The vector kernels need them because they accumulate in signed 16-bit lanes and their proxy undoes the offset — ProxyRwSse2/Avx2 <SplFmt_INT16>::S16 <CLIP_FLAG, SIGN_FLAG> XORs bit 15 on read/write, keyed on SB == 16 and DB == 16. The C++ kernel has no counterpart: it accumulates in a plain int and ProxyRwCpp is unsigned at both ends. So the constant only biased the sum, and write_clip clipped the negative result to 0.

The bug fires whenever that kernel runs with SB < 16, on either axis:

source	axis	kernel	unpatched
8-bit	horizontal	SB=16 DB=16	correct
8-bit	vertical	SB=8 DB=16	black
10-bit	either	SB=10 DB=16	black
12-bit	either	SB=12 DB=16	black
16-bit	either	SB=16 DB=16	correct

8-bit horizontal is the only case that escapes (that route converts to 16-bit before the scaler). On x86 none of it is reachable — the SSE2/AVX2 scalers replace the function in the Scaler constructor — so only macos-arm64 and linux-arm64 were affected.

Downstream this destroyed havsfunc's Bob(), which bobs fields with fmtc.resample(scalev=2, ...). QTGMC's Placebo and Very Slow (the only presets defaulting NoiseProcess=2, whose noise pass calls Bob()) came out ~+10/255 too bright; Draft (EdiMode='bob') came out nearly black.

Changes

• Scripts/patches/fmtconv-r31-arm-int-scaler.patch — applied by download-deps-{macos,linux}.sh right after the clone. Hard-fails the build if it ever stops applying, rather than silently shipping the bug back. The header carries the full analysis. Not submitted upstream (needs a GitLab account); upstream r31 does not fix this — building r31 reproduces it, as do yuygfgg's prebuilt arm64 binary and a local -O0 build.
• havsfunc patch 5 — makes Bob() resample at 16-bit. Redundant now and measured to give the same output, but it also covers the prebuilt Windows DLL and any build where the fmtconv patch is dropped. Applied on every platform so all platforms generate output from an identical havsfunc.
• fmtconv pinned to r31 everywhere. Upstream moved to GitLab in Aug 2023; the GitHub repo is an abandoned mirror whose last commit is literally "Repository moved to Gitlab", so macOS/Linux were cloning an unpinned stale post-r30 snapshot while Windows took the r30 zip. All three move together — r31 changed interlaced PAL-DV chroma placement, so a version skew would change chroma per platform.

Verification

Measured on interlaced_test.avi (source luma 130.896): Very Slow 140.794 → 130.895, Placebo 140.825 → 130.910, Draft 36.752 → 131.032, Slower 130.930 → 130.994 (control).

Correctness rather than just non-black: transposing the input, resampling along the other axis and transposing back gives output bit-identical to the direct result for luma in every format tested (8/10/12/16-bit, 4:2:0/4:2:2/4:4:4, up and down) and on every plane for 4:2:2 and 4:4:4.

Nothing here is exercised by the push gate (it downloads the published bundle rather than building it), so CI passing on this PR does not validate the change — that happens when deps-v1.7.0 is built and the app-side PR runs against it.