Add custom Ying memory profiler and prevent giant allocations

[velvia]

This change swaps out the explicit Jemalloc-based memory profiling code with my custom Ying profiler (https://github.com/velvia/ying-profiler), using the GlobalAlloc trait. Why?

1. This lets us get much richer and better Rust stack traces for the largest sources of retained memory from Rust code, esp async code, so we can pinpoint where the largest sources of allocation are happening, in a way that jeprof is not able to obtain;
2. This will allow us to prevent large memory allocations that is messing up devnet
3. Memory profiling is still using a sampling strategy, so it should remain light in CPU usage.

I have run local stress benchmark with the profiler turned on, and the QPS remained about the same as before (roughly ~360 when target = 400).

NOTE: the large memory allocation prevention will be updated in this PR shortly by uopdating the git pointer

This does not prevent one from using jemalloc, which can still be used by adding LD_PRELOAD and setting MALLOC_CONF etc. It does switch the default to the default Rust memory allocator underneath though.

[longbowlu]

It does switch the default to the default Rust memory allocator underneath though.

@mwtian it seems to collide with #5614 ?

[mwtian]

Looks good. This will disable process level jemalloc profiling in devnet. But depending on the Ying allocator is implemented, jemalloc profiling is probably still enabled in private testnet which would be fine.

[mwtian]

Is this removed to get clearer signal, or you have found incompatibilities? In the longer term we probably want to use jemalloc regardless of its profiling feature.

[velvia]

Yeah this is just so that we can do it environment by environment, to be safer