allocation profiler not threadsafe

[vilterp]

@vtjnash pointed out this thread safety bug not caught by our tests: #42768 (comment)

This can result in multiple threads corrupting the stack traces from each others' allocs, leading to a corrupted profile.

Fix involves keeping a buffer per thread.
EDIT: And/or maybe reusing the existing backtrace buffer in the ptls object. (See the linked thread, above)

cc @NHDaly

[Sacha0]

(Assigned to Pete at Pete's request :).)