Implement counters to track token usage across all LLM backends with model and backend labels

[psschwei]

Description:
Implement counters to track token usage across all LLM backends with model and backend labels.

Detailed Requirements:

1. Create counters in mellea/telemetry/metrics.py:
   • mellea.llm.tokens.input - Input/prompt tokens
   • mellea.llm.tokens.output - Output/completion tokens
   • mellea.llm.tokens.total - Total tokens (optional, can be computed)
2. Add attributes/labels:
   • gen_ai.system - Backend system (openai, ollama, etc.)
   • gen_ai.request.model - Model identifier
   • mellea.backend - Backend class name
3. Create helper function record_token_usage(input_tokens, output_tokens, model, backend)
4. Integrate with all backends that report token usage:
   • OpenAI: usage.prompt_tokens, usage.completion_tokens
   • Ollama: response.prompt_eval_count, response.eval_count
   • WatsonX: input_token_count, generated_token_count
   • LiteLLM: Via litellm's usage tracking
   • HuggingFace: Limited support (local models)

Files to Modify:

• mellea/telemetry/metrics.py - Add token counters
• mellea/telemetry/backend_instrumentation.py - Add metric recording alongside trace recording
• mellea/backends/openai.py - Record token metrics
• mellea/backends/ollama.py - Record token metrics
• mellea/backends/watsonx.py - Record token metrics
• mellea/backends/litellm.py - Record token metrics

Metric Definitions:

Metric	Type	Unit	Attributes
mellea.llm.tokens.input	Counter	{token}	model, backend, system
mellea.llm.tokens.output	Counter	{token}	model, backend, system

Acceptance Criteria:

• Token counters increment on each backend call
• Metrics have correct model/backend labels
• All backends with usage reporting integrated
• Zero overhead when metrics disabled