Prior to this commit, system prompt messages were included
by the `KurtVertexAI` as part of the `content` stream of messages.
However, VertexAI requires system messages to be specified separately.
This commit makes that change.

fix: system prompt formatting for Vertex AI

This will make it easy to work with LLMs in tests, without requiring any LLM connectivity/usage for
running existing tests whose inner LLM responses are already cached.

…xist

…mmitted one

feat: add `kurt-cache` package with `KurtCache` adapter

Prior to this commit, KurtCache would raise an unhandled YAML
serializing exception when storing a cache entry that included
structured data, because it was trying to serialize the Zod
schema object directly. This commit fixes that problem by
converting the schema to JSON schema (which is cleanly serializable)
before writing the YAML cache entry.

fix: KurtCache handling of schemas - serialize as JSON Schema

This maps to the new-ish `strict: true` feature of OpenAI which
enables constrained token sampling, but has certain caveats that
make it undesirable to turn on by default.

See issue #61 for more info.

feat: add `forceSchemaConstrainedTokens` to `KurtSamplingOptions`

This commit also updates the set of supported models to include the newer model snapshots. We also
update the tests to allow for specifying different models per test, and to mostly use the newer
models.

…ed-tokens

feat: add support for schema-constrained tokens in `KurtOpenAI`

…ema-constrained tokens case

Vertex AI is not yet capable of this feature, and we want to proactively warn the user that they
cannot expect a schema conformance guarantee.

… option

This ensures that changing the option will correctly cause a cache miss. But leaving it at the
default value of `false` won't change existing hashes because `false` is treated as "missing" in the
`mayHash` internal function.

…ined-tokens-error

fix: make KurtVertexAI properly throw KurtCapabilityError for the schema-constrained tokens case

…rained

fix: ensure `KurtCache` hashes the new `forceSchemaConstrainedTokens` option

…generateRawEvents`

Prior to this commit, that option was being silently dropped because it was neglected in the new
object that gets sent to the internal adapter. This commit introduces a usage of the spread operator
to ensure that no such options are dropped (while still allowing the replacement of other options).

fix: correctly pass through the `forceTool` option inside `KurtCache.generateRawEvents`