Currently there's no way to turn off embeddings caching for a single request once it's enabled globally via ai.caching.embeddings.cache. Passing cache: false is silently ignored, and ->cache(0) / ->cache(-1) still return an existing cached entry.

Now an explicit disable wins over the global config:

// global caching is on in config...

// before: still cached
str('hello world')->toEmbeddings(cache: false);
Embeddings::for(['hello'])->cache(0)->generate();

// after: both bypass the cache and hit the provider

Under the hood the request now tracks an explicit cache intent (default to config, or force on/off) instead of inferring it from whether a TTL was set, so a zero/negative TTL means "don't cache" rather than "cache with a 0s TTL". Tests added.

Fixes #691