docs(samples): backfill READMEs for anthropic + cohere + mistral provider samples

[mmercuri]

Summary

Backfills missing README.md for three provider samples on
feat/instrument-llm-providers (PR #94). Per the sample audit at
A:/tmp/adapter-samples-audit.md, anthropic, cohere, and mistral
all had working main.py files but no README. openai (CLEAN per
audit) is the canonical reference shape that the new READMEs match.

azure_openai and ollama are not present under samples/instrument/
on this branch (they live on feat/instrument-providers-azure-openai
and feat/instrument-providers-ollama respectively, per gh pr list),
so this PR does not touch them.

Per-provider features demonstrated (source-traced)

All claims trace to main.py for each sample and the corresponding
adapter at src/layerlens/instrument/adapters/providers/<name>_adapter.py.

anthropic

• Wraps client.messages.create and client.messages.stream via
  method substitution in AnthropicAdapter.connect_client.
• Sample exercises client.messages.create against
  claude-haiku-4-5-20251001.
• Emits model.invoke, cost.record, and (when tool_use blocks are
  present) tool.call. Anthropic's system kwarg is captured in
  request parameters; response.content is iterated as typed blocks.

cohere

• Wraps client.chat (v1), client.v2.chat (v2), and client.embed
  via method substitution in CohereAdapter.connect_client.
• Sample exercises Cohere v1 Chat (client.chat(...)) against
  command-r-plus.
• Token usage normalized from response.meta.billed_units; pricing
  falls back to pricing_unavailable: True for unknown models.

mistral

• Wraps client.chat.complete, client.chat.stream, and
  client.embeddings.create via method substitution in
  MistralAdapter.connect_client.
• Sample exercises client.chat.complete against
  mistral-small-latest using the native mistralai.Mistral
  client (not the OpenAI-compatible REST shim).
• Captures Mistral-native params: random_seed, response_format,
  safe_prompt, tool_choice.

PR #154 note

Each README carries a NOTE that the sample predates PR #154's
typed-events migration. Verified by reading
src/layerlens/instrument/adapters/providers/_base/provider.py on this
branch — _emit_model_invoke still calls self.emit_dict_event(...),
so dict-shaped events are still the current contract on
feat/instrument-llm-providers. PR #154 is open (not merged) and is
based off a different parent branch.

Test plan

• Sanity-render each README locally
• Cross-check every "what this demonstrates" bullet against
  main.py and the adapter source
• Confirm audit accuracy: anthropic, cohere, mistral each
  have main.py + __init__.py and no prior README.md on this
  branch via git ls-tree -r origin/feat/instrument-llm-providers
• Reviewer eyeball: provider-specific notes accurate (Cohere v1
  vs v2, Mistral native vs OpenAI-compatible, Claude version pin)

Out of scope

• Updating samples to use typed events (blocked on PR feat(instrument): typed events Bundle #6 (final) — _base + langchain + langgraph + agentforce + protocols #154 merge +
  rebase of this branch)
• Adding README for azure_openai / ollama (those samples don't
  exist on this branch)