Keypair auth does not use the shared session cache

Problem

Currently, keypair_credentials() has different behavior from other authenticators (externalbrowser, workload_identity):

For non-SPCS cases, keypair_credentials() returns the JWT directly as a Bearer token with X-Snowflake-Authorization-Token-Type: KEYPAIR_JWT. This seems to work in most cases, but:

1. We regenerates the JWT on every call (~10ms overhead per my benchmarking last year).
2. No participation in the session and keychain caching.
3. We're inconsistent with other auth methods, which all go through /login-request.

How the Python connector handles this

The Python connector always goes through /session/v1/login-request for keypair auth:

def update_body(self, body: dict[Any, Any]) -> None:
    body["data"]["AUTHENTICATOR"] = KEY_PAIR_AUTHENTICATOR  # "SNOWFLAKE_JWT"
    body["data"]["TOKEN"] = self._jwt_token

I can't find any evidence that they do the OAuth token exchange thing.

Proposed fix

Refactor keypair_credentials() to match the session caching pattern used by workload_identity_credentials() and externalbrowser_credentials().

We need to investigate whether SPCS also accepts session tokens from login_request or whether this needs to remain a separate path.

Start with keypair_credentials() and compare its flow with workload_identity_credentials() and externalbrowser_credentials(). Verify how non-SPCS keypair authentication currently handles the JWT, login_request, and session or keychain caching, then investigate whether SPCS accepts login_request session tokens. Done means keypair authentication follows the shared caching pattern where supported without breaking the separate SPCS path.