fix: Release full-quantize lookup on repeated setup

[carrerasdarren-cell]

Summary

• release the previous full-quantization lookup table before replacing it
• clear the member before allocation so cleanup remains safe if allocation fails

png_set_quantize() may be called more than once on the same png_struct, as
clarified while fixing #690. The earlier fix covers quantize_index in the
non-full-quantize branch, but the full-quantize branch still overwrites
palette_lookup on every call. With the default quantization bit widths, each
repeated call leaks 32,768 bytes.

The explicit NULL assignment also follows the allocation-failure cleanup
pattern used for the other png_struct members fixed in a22696b.

Verification

• A counting-allocator reproducer reports live_bytes=32768 after two calls on
  the unmodified branch and live_bytes=0 with this change.
• ctest --test-dir build-leak-proof --output-on-failure -j2 passes all 37
  tests.

[jbowler]

Same as #904

You are insisting on some etiolated idea of how to write code that does not comport with ANSI-C.

The data must be released when the png_struct (in this case) or png_info (in the other) is free'ed; in other words the data must be released when the app cleans up.

I suggest that submitted the same bug report multiple times is not productive.

[carrerasdarren-cell]

Closing this narrower follow-up to avoid duplicating the repeated-setter premise already under discussion in #904. I will keep any further evidence consolidated there.