fix: data race in ehash due to unsynchronized lazy init of Map.typ

[stuchl4n3k]

ehash uses double-checked locking to lazily initialize m.typ on first
use. The outer check reads m.typ without synchronization, while the
inner write happens under m.lock. This violates Go's memory model —
a concurrent reader can observe a partially written pointer.

Replace the bare *rtype field with atomic.Pointer[rtype] and use
Load/Store in ehash so the fast path is a single atomic load and the
write under the lock is correctly published to other goroutines.

[lrita]

Could you use unsafe.Pointer here, or upgrade the minimum dependency in go.mod to a version that supports this type?

[lrita]

We can move this to the beginning of the function so it only loads once for the most common scenarios. Meanwhile, let's refine the double-check logic to make it works fine.