Bug fixes (critical):
- callbacks.py: add ABC to UiPathDynamicHeadersCallback so @AbstractMethod
  is enforced — previously the class could be instantiated without implementing
  get_headers(), silently deferring the error to runtime
- clients/anthropic/chat_models.py: make vendor_type a required field (no default)
  and seed api_config with VendorType.VERTEXAI as placeholder; the old default of
  VendorType.ANTHROPIC caused the model validator to immediately raise ValueError,
  making the class impossible to construct directly; also removed dead
  VendorType.ANTHROPIC / VendorType.AZURE branches from _anthropic_client and
  _async_anthropic_client that could never be reached, and unused imports
- clients/bedrock/utils.py: fix json.loads(kwargs.get('body', {})) — passing a
  dict as default caused TypeError since json.loads requires str/bytes; changed
  default to '{}' string; also corrected converse() and converse_stream()
  system parameter type annotation from str|None to list[dict[str,Any]]|None to
  match the list[dict] value langchain-aws actually passes
- clients/normalized/chat_models.py: add 'stream': True to request body in
  _uipath_stream and _uipath_astream — without it the server returns a regular
  non-streaming JSON response instead of SSE; fix fragile SSE prefix stripping
  from chunk.split('data:')[1] (splits on all occurrences) to
  chunk[len('data:'):].strip(); fix _generate_chunk to set text=content or ''
  instead of text=original_message (raw JSON wire data)
- factory.py: replace bare model_info['vendor'] key access in get_embedding_model
  with model_info.get('vendor') + explicit ValueError, matching the safe pattern
  already used in get_chat_model

Bug fixes (minor):
- base_client.py: correct return type annotations on _astream, _uipath_astream,
  and uipath_astream from AsyncIterator to AsyncGenerator (these are async
  generator functions that contain yield, not plain iterables)
- demo.py: replace unsafe eval(expression) with an AST-based arithmetic-only
  evaluator to prevent arbitrary code execution via LLM-generated input

Version bumps:
- packages/uipath_langchain_client: 1.5.10 -> 1.6.0
- src/uipath/llm_client:             1.5.10 -> 1.6.0

Bug #5 — UiPathFireworksEmbeddings: self.model returned wrong model name
  Without an explicit field override, Pydantic kept two separate fields:
  FireworksEmbeddings.model (default 'nomic-ai/nomic-embed-text-v1.5') and
  UiPathBaseLLMClient.model_name (populated by the caller via alias 'model').
  self.model therefore always resolved to the stale FireworksEmbeddings default,
  silently ignoring whatever model name the caller passed.
  Fix: add model: str = Field(default='', alias='model_name') — the same override
  pattern used by every other embedding class — to collapse the two fields into
  one. Replace self.model with self.model_name in all four embed methods.

Bug #8 — UiPathEmbeddings: silent HTTP error swallowing + missing model field
  Two issues in embed_documents / aembed_documents:
  1. uipath_request / uipath_arequest were called without raise_status_error=True,
     so HTTP 4xx/5xx responses were silently accepted. The subsequent
     response.json()['data'] access would then raise a confusing KeyError or
     AttributeError rather than a clear HTTP error.
  2. The request body only contained {'input': texts}. The normalized API
     requires a 'model' field for routing, matching the pattern already used
     by the normalized chat model (_default_params includes 'model').
  Fix: pass raise_status_error=True and add 'model': self.model_name to the
  request body in both the sync and async paths.

The normalized routing layer resolves the model from the URL/connection
context — the caller does not need to send it in the JSON body.

- normalized/chat_models.py: remove 'model': self.model_name from _default_params
- normalized/embeddings.py: remove 'model': self.model_name from both
  embed_documents and aembed_documents request bodies

- Restore VendorType.ANTHROPIC as default for both api_config.vendor_type
  and the vendor_type field
- Restore all four vendor branches (ANTHROPIC, AZURE, VERTEXAI, AWSBEDROCK)
  in _anthropic_client and _async_anthropic_client
- Restore original imports (Anthropic, AnthropicFoundry, AsyncAnthropic,
  AsyncAnthropicFoundry) that were removed in the earlier fix
- Remove the long explanatory comment on api_config
- Keep the original error messages verbatim

- clients/vertexai/chat_models.py: pass dict(self.uipath_sync_client.headers)
  and dict(self.uipath_async_client.headers) instead of the raw httpx.Headers
  objects, matching the pattern used in clients/anthropic/chat_models.py
- demo.py: import RoutingMode from the public uipath_langchain_client.settings
  API instead of the internal uipath.llm_client.settings.constants path;
  merge with the existing get_default_client_settings import

These vendors are not yet supported by UiPath's gateway, so no api_flavor
routing header should be emitted for them.