feat: allow per-request spread via SpreadCtxKey context override

[ecordell]

Problem

Spread is a single value in the balancer config, so every key gets the same number of candidate backends. A caller that detects a hot key cannot spread only that key across more backends. Raising the global spread dilutes cache locality for every key to protect a few.

Change

A request can carry a uint8 value at SpreadCtxKey in its context. The picker then uses that spread instead of the spread from the balancer config. The picker ignores a value of zero or of any other type.

Also, this aligns the not-enough-members fallback with its comment:

• A spread that exceeds the count of ready members now uses all ready members. Before, the pick collapsed to a single member.
• A pick against an empty ring queues the RPC. The clamp needs an explicit guard for this case, because FindN(key, 0) returns an empty result and no error.

Tests

Four new tests: the override picks as the equivalent config spread would, a zero override uses the config spread, a spread above the member count uses all members, and a pick against an empty ring queues instead of panicking. The empty-ring test failed with a panic before the guard. The suite passes with -race.