Fix out-of-bounds read in motion compensation (plm_video_process_macroblock)

[hardw00t]

Summary

plm_video_process_macroblock() bounds-checks only the base source/destination
indices si/di against max_address, but the half-pel prediction cases dereference
s[si + 1], s[si + dw] and s[si + dw + 1], and PLM_BLOCK_SET iterates a full
block_size × block_size region. The deepest source byte actually read is:

si + (block_size-1)*dw + (block_size-1) + (odd_v ? dw : 0) + (odd_h ? 1 : 0)

For si near max_address with an odd (half-pel) motion component this reaches up to one
row (dw) past the end of the plane. Since all three frames share a single
PLM_MALLOC(frame_data_size * 3) chunk with no padding, when the reference plane is the last
plane the read runs off the end of the allocation — an out-of-bounds read on crafted input.
Motion vectors from plm_video_decode_motion_vector() are wrapped to ±(fscale<<4) but not
clamped to the frame geometry, so a malformed stream can place the prediction source at the edge.

Destination writes (d[di]) stay in bounds, so this is a read-only overflow — impact is a
potential crash (read into an unmapped page) and, since the out-of-bounds bytes are averaged
into decoded pixels, a limited information-leak into the output frame.

Fix

Bound the actual deepest source and destination offsets against the plane size, not just the
base indices:

unsigned int plane_size = dw * (self->mb_height * block_size);
unsigned int max_di = di + (block_size - 1) * dw + (block_size - 1);
unsigned int max_si = si + (block_size - 1) * dw + (block_size - 1)
    + (odd_v ? dw : 0) + (odd_h ? 1 : 0);
if (si >= plane_size || di >= plane_size || max_si >= plane_size || max_di >= plane_size) {
    return; // corrupt video
}

Testing

• Found via libFuzzer + AddressSanitizer fuzzing of the decode path (plm_create_with_memory
  → plm_decode_video).
• With this change, the crafted stream that triggered the ASan heap-buffer-overflow in
  plm_video_process_macroblock no longer overflows (the macroblock is treated as corrupt and
  skipped), and valid MPEG1 streams still decode identically.
• Header compiles cleanly with -Wall -Wextra.

I have a minimal reproducer that I'm happy to share privately if useful — just let me know the
best channel; I kept it out of this public PR.