fix: handle nil bucket_id in storage get/update/delete operations

[p0rtale]

Starting from crud 1.7.0, storage asserts if opts.bucket_id is nil during get, update, or delete operations. Since older routers (< 1.7.0) do not pass bucket_id, this causes protocol incompatibility.

This PR adds a nil check to skip bucket ref when bucket_id is missing, restoring backward compatibility with old routers.

It also introduces a rate-limited warning and the tnt_crud_storage_nil_bucket_id_compat_total metric to track
this state.

• Tests
• Changelog
• Documentation

Closes TNTP-7622

[vakhov]

Thanks for the PR. I have suggestions here:

• the workaround itself looks reasonable as a temporary solution for rolling cluster upgrades, but it should be made observable. Right now, when bucket_id == nil, storage silently skips bucket_ref, so during a partial upgrade some traffic goes through a path with weaker guarantees, and operations won’t see it.
• maybe it’s worth adding a rate-limited warning through a shared helper, for example warn_nil_bucket_id_compat_once(operation, space_name, space.engine), to avoid duplicating the same logic across get/update/delete. The message could be something like: crud.%s_on_storage called without bucket_id; old router compatibility mode is used, bucket_ref is skipped, rebalance safety is reduced until routers are upgraded; space=%s engine=%s.
• it may also be worth adding a storage-side counter: tnt_crud_storage_nil_bucket_id_compat_total with labels like operation, engine, and possibly safe_mode.
• also, it may be worth documenting the upgrade workaround explicitly for clusters upgrading from < 1.7.0 to >= 1.7.0. The docs should say that during a rolling upgrade storages should be upgraded first, and routers afterwards. While old routers are still running, storages may enter the compatibility path where bucket_id is missing and bucket_ref is skipped, so should expect the warning/counter above until all routers are upgraded.

[ita-sammann]

It bothers me that where we have just added safe mode in 1.7.0 and now we are already punching holes in it. AFAIK 1.7.x router can work with 1.6.x storages. So user can just update routers first and update storages after that. And we don't need to add this check for nil bucket_id.

I suggest that we add instruction how to update cluster to the newer crud versions to README.md:

• Update storages from pre-1.6.0 (role-based model) to 1.6.1
• Update routers from pre-1.6.0 to 1.7.4 (safe mode)
• Update storages from 1.6.1 to 1.7.4