feat(api): expose memory_type on create and update

[doctatortot]

Summary

POST /v1/memories hardcoded MemoryType::Pinned as the default for direct content creation, and PUT /v1/memories/{id} had no way to change the type at all. As a result every API-created memory ended up pinned, and there was no recovery path short of delete-and-recreate.

This PR:

• Adds memory_type: Option<String> to CreateMemoryBody. When provided, the value is parsed via MemoryType::FromStr (returns Validation / 400 on unknown variants). When absent, the existing default of Pinned is preserved so this is not a breaking change.
• Adds memory_type: Option<String> to UpdateMemoryBody with the same parse-and-validate flow.

Motivation

I ran into this myself. After migrating ~170 memories into my omem instance via the API, every single one ended up memory_type: pinned and got the 1.5× pinned_boost in stage_rrf_fusion — effectively a uniform inflation across the corpus that defeats the boost's purpose. There was no API path to demote them; the only fix was delete-and-recreate.

Tests

• test_create_memory_with_type: POST with memory_type=insight returns insight; POST without it still returns pinned (default-preserved, no regression for existing callers).
• test_update_memory_type: PUT memory_type=insight on an existing pinned memory demotes it.
• test_update_memory_type_invalid: PUT with an unknown variant returns 400 Bad Request rather than silently accepting.

Note for reviewers

The create-side default of Pinned is preserved here intentionally to avoid a behavior change in a single PR. Whether Insight would be a better default is worth a separate design discussion — happy to open a follow-up issue if you'd like to push on it.

CI on this PR will still surface the pre-existing cargo fmt --check debt and the rustls 0.23 CryptoProvider test failures (the latter is what #6 addresses); my new tests rely on the same fix #6 applies, so once #6 lands they'll go green. I validated them locally on my fork with the #6 patch cherry-picked in: all four scoped test buckets pass (store 34/34, retrieve::pipeline 19/19, the new create test 1/1, the new update tests 2/2).

[yhyyz]

Applied in 628b757 — thanks @doctatortot!

Clean addition — the memory_type field on create/update fills a real gap (bulk imports all ending up as pinned was a pain point we'd heard before). All 3 new tests pass. Default preserved for backwards compat.

Re: the design note about whether Insight should be the default instead of Pinned — worth discussing in a separate issue if you want to push on it.