feat: add multi-provider NLP support for intent parsing

[Noooblien]

Currently the parser relies on pattern matching and heuristics. Add support for multiple LLM providers as the NLP backbone so teams can choose based on cost, latency, or preference.

Providers to support:

• anthropic — Claude API (claude-haiku for speed, claude-sonnet for accuracy)
• openai — GPT-4o / GPT-4o-mini
• google — Gemini 1.5 Flash
• mistral — Mistral Small (good for cost-sensitive deployments)
• cohere — Command R (strong on structured extraction tasks)

Proposed interface:

const parser = new IntentParser({
  provider: "anthropic", // swap to "openai" | "google" | "mistral" | "cohere"
  apiKey: process.env.API_KEY,
  model: "claude-haiku-4-5" // optional override
})

What this unlocks:

• Teams not on Anthropic can still use intent-parser
• Benchmark accuracy and latency across providers per corridor
• Fallback to a secondary provider if primary is down
• Cost optimization — route simple intents to cheaper models

Suggested approach:

• Abstract LLM calls behind a ProviderAdapter interface
• Each provider gets its own adapter — AnthropicAdapter, OpenAIAdapter etc.
• Keep regex heuristics as the fast path, LLM as the enrichment layer
• Add a parsed_by field in response showing which provider and model was used

[kritarth1107]

Picking this up now. Starting with the provider adapter pattern so each LLM sits behind a common interface — makes swapping clean without touching core parsing logic.

Will begin with Anthropic and OpenAI first, then roll in Gemini, Mistral, and Cohere once the abstraction is solid. Will update here as PRs go up.