Update dtoa to version from Ruby 4.0

[jhawthorn]

Fixes #527

Upstream commits:

• 3acc81d9e4 Fixed race in dtoa [Bug #17612]
• d0a822eec5 Fix dtoa buffer overrun
• bc8cc68aef Make dtoa.c buildable alone
• d17edf3a17 [Bug #20705] Update strtod implementation
• b3b1d7b295 Remove spinlock in dtoa's Balloc, use xmalloc
• 0887d82406 dtoa.c: Add shortcut if arguments are zero
• d0c966adcd dtoa.c: Extract macro to update 5powers Bigint cache atomically
• f047174cf1 dtoa.c: Check memory allocation failures
• 624538ba92 dtoa.c: Check integer underflow
• 44693ee329 Fix a possible memory leak in dtoa

[byroot]

Shouldn't we fix dtoa to use xmalloc instead?

[jhawthorn]

I'm not sure there's value in doing so. These allocations are short-lived and running the GC doesn't help us remove them (they aren't associated with a Ruby heap object) so I'm not sure it helps for them to contribute to GC pressure.

This was changed upstream in ruby/ruby@bc8cc68

[byroot]

Sorry, I should have explained my train of thoughts. I agree there's no GC benefits, especially since when you use xfree (instead of the new xfree_sized) this likely cause a malloc_usable_size call which make the whole thing slower.

However, I recently ran into a bunch of bugs related to using xfree with malloc and vice-versa, so I tend to think it's simpler to just always use the Ruby allocator.

But really up to you.

Anyway, as you said bgidecimal should probably move away from vendoring dtoa.

[tompng]

I'd like to use malloc and free instead of xmalloc and xfree in bigdecimal's dtoa.
dtoa and strtod may call MALLOC several times, and if the second call fails to allocate, dtoa frees the first allocated one.

b1 = Balloc(b->k+1);
if (!b1) {
    Bfree(b);
    return NULL;
}

bd = Balloc(bd0->k);
if (!bd) goto retfree;

If we use ruby_xmalloc, and the second allocation raises NoMemoryError, this mechanism will fail, and will cause memory leak.
So I think current dtoa.c in ruby/ruby can be considered designed to be use with malloc and free

move away from vendoring dtoa.

Agree