feat: make inference provider optional

Inference provider implementations behave inconsistently—some can load
normally without an API key (even though it’s required for proper
functionality), while others simply fail. To ensure consistency, we
should handle inference providers the same way we do vector I/O
providers: treat most of them as optional, with vLLM being the
exception.

Relates to: RHAIENG-1178
Signed-off-by: Sébastien Han <seb@redhat.com>

Author: leseb

distribution/run.yaml

@@ -20,7 +20,7 @@ providers:
       max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
       api_token: ${env.VLLM_API_TOKEN:=fake}
       tls_verify: ${env.VLLM_TLS_VERIFY:=true}
-  - provider_id: bedrock-inference
+  - provider_id: ${env.AWS_ACCESS_KEY_ID:+bedrock}
     provider_type: remote::bedrock
     config:
       aws_access_key_id: ${env.AWS_ACCESS_KEY_ID:=}
@@ -36,7 +36,7 @@ providers:
   - provider_id: sentence-transformers
     provider_type: inline::sentence-transformers
     config: {}
-  - provider_id: watsonx
+  - provider_id: ${env.WATSONX_API_KEY:+watsonx}
     provider_type: remote::watsonx
     config:
       url: ${env.WATSONX_BASE_URL:=https://us-south.ml.cloud.ibm.com}