Fix GC allocation rate underflow when the heap shrinks between cycles

[gemtool]

The GC measures heap growth since the end of the last incremental cycle as atomicstarttotalsizebytes - endtotalsizebytes (and totalbytes - endtotalsizebytes in luaC_allocationrate). Both operands are size_t, so when the heap is smaller than it was at the end of the last cycle, the subtraction wraps around. This can happen in two ways:

• a full collection (lua_gc(L, LUA_GCCOLLECT, 0)) frees memory but doesn't update the incremental cycle statistics
• thread stacks are shrunk during marking, for example after a coroutine that recursed deeply returns

	before	after
lua_allocationrate() after a full collection	-9223372036854775808	0
heap growth seen by the pacer after a coroutine stack shrink (646KB -> 348KB)	18446744073709253349 bytes	0 bytes
allocation rate used by the pacer after a full collection	1.08e21 bytes/s	0 bytes/s

The double to int64 conversions of these values are undefined behavior (reported by UBSan in luaC_allocationrate and getheaptrigger). lua_allocationrate was added in 0.731 to pace the GC externally, so an embedder that runs a full collection gets a garbage rate right after it.

This change treats a heap that shrank as zero growth. When the heap grew, the results are unchanged. The fix is behind DFFlag::LuauGcHeapShrinkFix and the flag-off paths keep the original expressions.

Resetting the cycle statistics in luaC_fullgc would only cover the first case, so the growth is clamped instead.

Added ApiAllocationRateAfterFullGC, which fails without the fix (lua_allocationrate returns -9223372036854775808).

[tommyscholly]

Please reset gcstats.endtimestamp and gcstats.endtotalsizebytes at the end of luaC_fullgc() as well.
Otherwise, lua_allocationrate() will continue to report 0 after a full collection until the heap exceeds its previous size, even when allocations occur.
Additionally, I would like to see this test allocate after the full GC and check that lua_allocationrate() becomes positive.

Additionally additionally we should probably flag these changes I'm requesting as well.

[gemtool]

Thanks, good catch! luaC_fullgc now resets gcstats.endtimestamp and gcstats.endtotalsizebytes, behind the same LuauGcHeapShrinkFix flag. The test now allocates after the full collection and checks that lua_allocationrate becomes positive.