Loader.Load allocates a batchRequest + channel before the cache check

[Zac300]

Loader.Load allocates the batchRequest (and its done channel) at the very top of the function, before the cache lookup:

func (l *Loader[K, V]) Load(originalContext context.Context, key K) Thunk[V] {
	ctx, finish := l.tracer.TraceLoad(originalContext, key)
	req := &batchRequest[K, V]{
		key:  key,
		done: make(chan struct{}),
	}
	...
	if v, ok := l.cache.Get(ctx, key); ok {
		...
		return v   // req is never used on this path
	}

req is only used on a cache miss — it's sent to the batcher (l.curBatcher.input <- req) and read by the returned thunk. On a cache hit it's allocated and immediately discarded, so every cached Load pays for a struct + channel allocation it never uses. In read-heavy workloads where the same keys are loaded repeatedly, this is pure waste and shows up prominently in allocation profiles.

Moving the allocation below the cache-hit early-return eliminates it for all cache hits.

Benchmark (Load on a primed/cached key, -benchmem):

	allocs/op	B/op	ns/op
before	3	160	45.7
after	1	16	20.5

The remaining allocation on the hit path is the no-op finish closure returned by NoopTracer.TraceLoad (a fresh func(Thunk[V]){} per call, which isn't free inside a generic function) — happy to file that separately if of interest.

The full package test suite passes with the change. Fix proposed in #123.

[pavelnikolov]

Thank you for your contribution! Optimizations and improvements are always welcome!