Speed up TRefCountedTracker allocation accounting

Every `TRefCounted` allocation/free bumps a per-thread counter in `TRefCountedTracker`. That hot path went through two `Y_NO_INLINE` TLS accessors (one per thread-local) plus a facade call. This reworks it:

1. Merge the three per-thread thread-locals (`...Slots`, `...SlotsBegin`, `...SlotsSize`) into one `constinit thread_local` struct -- one cache line, one TLS base, no dynamic-init guard.
2. Read it directly instead of through the generated `Y_NO_INLINE` accessor, collapsing the per-call accessor `call` into a couple of `%fs`-relative instructions.
3. Keep it fiber-safe by construction: the counter entry points are now `private` and `friend`ed only to `TRefCountedTrackerFacade`, the single cross-library boundary, which is pinned `YT_PREVENT_TLS_CACHING`. The hot path inlines into that out-of-line function, so the thread pointer is re-read on every call and can never be cached across a fiber migration. Every other function that touches the TLS (`*Slow`, `GetLocalSlot`, the thread-exit reclaimer) is pinned the same way.

Isolated microbenchmark (`BM_RefCountedTracker_AllocateFreeInstance`, added here), pinned core, allocate+free pair:

| | ns/pair |
|---|---|
| before | 11.5 |
| after | 4.4 |

`AllocateInstance` body: 8 instructions, 0 internal calls, no jmp.
commit_hash:8601e1eaa2ee25e3e43a8792fb10b4c02c1b2cef

0 files changed, 0 insertions, 0 deletions