Fix three signed-overflow/UB sites in the RAS decoder

[iliasabk]

Summary

Fixes #413 — three UBSan sites reachable from jas_image_decode on crafted RAS files:

1. ras_dec.c:352 — the bit accumulator z is int; (z << 8) overflows into the sign bit for depth >= 24. z is now uint_fast32_t (consumers mask with RAS_ONES anyway).

2. ras_dec.c:206 — hdr.length = RAS_ROWSIZE(&hdr) * hdr.height overflows int_fast32_t (32-bit on i386/Darwin) for large width * depth * height, e.g. width=8192, depth=32, height=65536 → 32768 * 65536 = 2^31. RAS_ROWSIZE now computes in int_fast64_t, as does the pad computation which also multiplies width * depth.

3. ras_dec.c:412 — numcolors = 1 << hdr->depth is UB for depth = 32 (checked only after the UB). The bound check now runs on JAS_POW2_X(uint_fast64_t, hdr->depth) before numcolors is materialized; depth <= 8 is required anyway since RAS_CMAP_MAXSIZ is 256.

Testing

No PoC was attached to the issue, so I generated minimal reproducers (32-byte RAS headers + data bytes):

• depth=32, maptype=EQUALRGB → ras_dec.c:412:17: runtime error: shift exponent 32 is too large
• type=OLD, 8192x65536x32 → ras_dec.c:206:34: runtime error: signed integer overflow: 32768 * 65536
• depth=32, data byte 0x80 → ras_dec.c:352:12: runtime error: left shift of 8388608 by 8

All three report cleanly after the patch; malformed files are rejected gracefully.