perf: cut per-frame CPU cost in the render path

[gridboy]

What this changes

Removes the dominant CPU cost from the real-time render loop, and fixes the bugs and leaks found in the export paths along the way.

Why

Three hot spots, all per frame:

	before	after
depth clamp	4.69 ms	0.71 ms	6.6x
mesh index build (detail 1)	0.087 ms	0.019 ms	4.6x
malloc/free traffic	1.5 MB	none	recycled

The depth clamp called [ctrl getDetail], [ctrl getMin] and [ctrl getMax] inside the per-pixel loop — up to 7 objc_msgSend per pixel — plus two fmod() calls, over 307200 pixels, every frame. The accessors are hoisted and the modulo replaced by the loop's own column index. The export paths in AppDelegate.m already used the ivars directly, so this brings the render path in line with what the author wrote elsewhere.

The mesh index builder walked x outer / y inner, striding FREENECT_FRAME_W * 2 bytes per step through a row-major buffer. Loop order swapped so the walk is sequential.

Frame buffers were freed and re-malloced every frame at 600 KB and 900 KB — above the allocator's mmap threshold, so each frame paid two mmap/munmap round trips plus page faults. -recycleDepthData: / -recycleVideoData: hand them back for reuse; calling free() on those buffers stays correct, so no existing caller was invalidated.

Also fixed here:

• saveSTLB counted facets in an NSUInteger * — a pointer — so faceCount + 2 was pointer arithmetic advancing by 16. The binary STL header announced 8x more facets than were written.
• binaryVector: allocated NSMutableData without calling init.
• All three export paths leaked their 600 KB depth buffer; savePly and saveSTL also leaked their accumulator strings.
• PLY/STL text built one NSString per coordinate — 12 temporaries per facet. Now a single format pass, byte-identical output.

How it was verified

• sh scripts/syntax-check.sh passes
• Behaviour-preserving changes backed by a comparison against the original: both rewritten loops were run against the originals over 960 randomised parameter combinations (random depth buffers, detail 0-40, min/max sweeps). Byte-identical output buffers; identical index sets and counts for the mesh.
• Performance claims come with numbers: measured with a benchmark carrying real objc_msgSend traffic on an M-series Mac, 200 frames per configuration. First run reported the mesh loop at 0.000 ms because the compiler had eliminated it as dead — the numbers above are from the corrected run with the result consumed.
• Built in Xcode
• Run against real Kinect hardware

The last two cannot be done here: Xcode 3 project format, 10.6 SDK, 2010 libusb binary, and no Kinect. Reviewers with the hardware should treat the render output as unverified.

Provenance

• No third-party licence header, copyright notice, CONTRIB or AUTHORS file altered
• No new vendored code