avoid signed overflow in cache memory limit

[metsw24-max]

sharp.cache({ memory }) passes the megabyte value through to the native cache helper, which reads it as a 32-bit int and multiplies by 1048576 to convert to bytes before handing the result to vips_cache_set_max_mem. The multiply runs in 32-bit signed arithmetic and only widens to the size_t the function expects afterwards, so any value of 2048 or above overflows first. A 2048MB request wraps to a negative number that sign-extends to roughly 16 exabytes, which leaves the operation cache effectively unbounded, while exactly 4096MB wraps to zero and silently disables it. I came across this while confirming that larger cache limits were honoured and saw the reported memory.max had no relation to the value passed in. Casting the operand to size_t before the multiply keeps the conversion in 64-bit so the byte count stays correct across the documented range. I have limited the change to that one expression and added a regression test at 4096MB, which reported a maximum of zero beforehand.

[lovell]

Thanks for the PR. Whilst we're fixing this, I think we might also want to use Uint32Value instead of Int32Value for the cache values as they should all be positive integers. I guess we could also verify these values in the JavaScript layer.

[metsw24-max]

Makes sense. Switched all three cache values to Uint32Value. I kept the size_t widening on memory because Uint32Value on its own still does the MB-to-bytes multiply in 32-bit and wraps from 4096MB up.

Also added a guard in lib/utility.mjs that rejects any cache value that isn't a non-negative integer, with a test for it. All squashed into the one commit.

[lovell]

Thanks for the updates, it looks like we might have to skip the max memory test on 32-bit systems. The buildPlatformArch() function should provide a suitable value for this.

[metsw24-max]

Good shout. The 4096MB byte count overflows a 32-bit size_t too, so I now skip that test when buildPlatformArch() reports a 32-bit arch (arm, ia32, wasm32). Runs as before on 64-bit and skips cleanly when I force npm_config_arch=ia32.

[lovell]

Thank you. (The wasm32 test failure is unrelated, I'm fixing this now.)