stream: group bool fields at tail of serverStream to eliminate false sharing with mu - #9361

Reorder `serverStream` fields so both bool fields are grouped at the tail of the struct, and move `mu` to a separate cache line from the tail bools.

Why

`serverStream` had 2 `bool` fields placed between aligned fields, each forcing alignment padding before the next field:

Total: 8 B wasted. Grouping both bools at the tail eliminates the interior padding, reducing `serverStream` from 256 B → 248 B, saving 8 B per server-side RPC stream unconditionally.

Additionally, `serverHeaderBinlogged` is written unsynchronised in `SendHeader` and `Send` (it is documented as not needing synchronization). In the tail-grouped layout, the bools land at offset 240+ — which was the same cache line as `mu` (offset 232, cache line 3: 192–255). Concurrent writes to `serverHeaderBinlogged` invalidate the cache line before every `mu.Lock()`, adding 5–21× latency overhead under concurrent load.

Moving `mu` to immediately before `trInfo` (which it guards) places it at offset 184 (cache line 2: 128–191). The tail bools remain at offset 240+ (cache line 3), so the false-sharing is eliminated. No size change from this step — `serverStream` stays at 248 B.

See benchmarks in #9349.

Benchmark — `mu.Lock()` latency with stressor goroutines writing `serverHeaderBinlogged` concurrently:

Code clarity

Both bools are unguarded. The bool group uses a single `// Not guarded by mu` section, preserving the existing per-field comments. `mu` is placed immediately before `trInfo` (which it guards) with a comment explaining the cache-line placement.

Closes #9349

RELEASE NOTES:

• core: reorder `serverStream` fields to eliminate alignment padding (saving 8 bytes per server-side RPC stream) and cache-line false sharing between `serverHeaderBinlogged` and `mu` (up to 21× `mu.Lock()` speedup under concurrent load).