Micro-optimize TCallback propagating storage handling

Adds Google Benchmark microbenchmarks for the `TCallback` / `TPropagatingStorage` hot paths under `yt/yt/core/actions/benchmarks` (a `TTraceContext` is constructed at startup so the realistic production switch handler is active in every benchmark), then optimizes the path so that `TCallback::Run()` pays less per invocation.

## Headline benchmark deltas

Release build, single thread, trace-context switch handler registered:

| Benchmark                            | Trunk  | Opt    | Δ     |
|--------------------------------------|--------|--------|-------|
| `BM_Capture_NoStorage`               |  34 ns |  31 ns |  -7%  |
| `BM_Capture_EmptyStorage`            |  38 ns |  35 ns |  -6%  |
| `BM_Capture_NonEmptyStorage`         |  38 ns |  36 ns |  -5%  |
| `BM_Capture_NoPropagate_WithStorage` |  24 ns |  26 ns |  noise |
| `BM_Run_NullPS`                      |  30 ns |  16 ns | **-46%** |
| `BM_Run_EmptyPS`                     |  74 ns |  44 ns | **-41%** |
| `BM_Run_NonEmptyPS`                  |  76 ns |  44 ns | **-41%** |
| `BM_Run_NoPropagate`                 |   2 ns |   2 ns | unchanged |
| `BM_CaptureAndRun_NullPS`            |  67 ns |  52 ns | **-23%** |
| `BM_CaptureAndRun_NonEmptyPS`        | 112 ns |  79 ns | **-30%** |

`BIND_NO_PROPAGATE` is unchanged.

## Changes vs trunk

### Manager dispatch

- `TPropagatingStorageManager` loses its `LeakySingleton` wrapper and becomes a `constinit` static in `NDetail`. The per-call double-checked-locking pointer load on every `TCallback::Run` goes away; access is now a plain address.
- `TPropagatingStorageManager::SwitchPropagatingStorage` stays out-of-line in the cpp to avoid bloating every `TCallback::Run` instantiation.
- `TPropagatingStorageGuard`'s ctor and dtor move from the cpp into the inl, so the install/restore call site goes directly to `PropagatingStorageManager.SwitchPropagatingStorage(...)` without the trunk's `TPropagatingStorageManager::Get()->...` member-pointer indirection.

### FLS pointer caching

- `TPropagatingStorageGuard` gains a `TFls*` cache populated on its first `SwitchPropagatingStorage` call and reused on its second. Drops one `Y_NO_INLINE` `GetCurrentFls()` call per `Run`.
- Inside `SwitchPropagatingStorage`, the `TFls*&` parameter is hoisted once into a local register at function entry so subsequent in-function uses don't go through the reference.

### Storage internals

- `TPropagatingStorage::Impl_` moves from `TIntrusivePtr<TImpl>` (private nested) to `TIntrusivePtr<TPropagatingStorageImpl>` (`NDetail`, defined in `propagating_storage-inl.h`). This lets the read paths — `IsNull`, `IsEmpty`, `FindRaw`, and the templated `Find` / `Has` — all inline at callers. Mutating ops (`ExchangeRaw`/`RemoveRaw`/`Clone`) stay out-of-line.
- `IsNull`, `IsEmpty`, `FindRaw`, `CurrentPropagatingStorage`, and `GetCurrentPropagatingStorage` move from cpp out-of-line definitions to inl inline definitions accordingly.
- `TPropagatingStorage`'s explicit special-member declarations are dropped (relying on implicit generation), except for the default ctor (`= default`) which has to stay declared because the private converting ctor would otherwise suppress it.
- `TPropagatingStorageImpl`'s copy ctor is hand-written (members init explicitly) because `TRefCounted`'s copy ctor is deleted; it preserves trunk's behavior that `Clone()` carries the signal subscriber lists alongside the data map.

### Trace-context lookup

- `TryGetTraceContextFromPropagatingStorage` is now defined inline in `trace_context-inl.h` (one definition; the previous `Fast` variant is merged in). Hot-path callers — `OnPropagatingStorageAfterSwitch` and the `YT_ASSERT` in `OnPropagatingStorageBeforeSwitch` — inline the lookup. An out-of-line copy is force-emitted via `[[gnu::used]]` so the GDB fiber printer (`devtools/gdb/yt_fibers_printer.py`) can still resolve the symbol at runtime.

### Inlining nudges

- `TPropagateMixin::MakePropagatingStorageGuard` gains `Y_FORCE_INLINE` so it actually inlines into `TBindState::Run` rather than being emitted as a separate weak symbol per instantiation.
- `TPropagatingStorage::IsNull` is also marked `[[gnu::used]]` for the same GDB-symbol-resolution reason as `TryGetTraceContextFromPropagatingStorage`.
commit_hash:86793cebfc5ff891affab25dcc67d987e553854c

0 files changed, 0 insertions, 0 deletions