fix(rollup): preserve provider through the rollup index

[bravo1goingdark]

Why

`/v1/rollups` and `/v1/stats?group_by=model` were returning `provider=""` for every row, even though events were ingested with `provider="mistral"` / `"groq"` / etc. Anything downstream that grouped by provider collapsed onto a single empty bucket — Obol's BY PROVIDER chart was the visible victim.

Two-part root cause:

1. `EngineConfig.rollup_dims` was `[USER_ID, API_KEY_ID, MODEL]`. Provider wasn't a grouping dim of the keplordb rollup index, so the in-memory aggregation merged across providers without distinguishing them.
2. `RollupKey` in `kdb_store.rs` only had four fields (day + 3 dims) and `RollupRow.provider` was hard-coded to `String::new()` in the construction path.

Fix

• `rollup_dims = [USER_ID, API_KEY_ID, MODEL, PROVIDER]` — provider becomes a grouping dim.
• `RollupKey` gains a `provider` field, populated from `key.dims.get(3)`.
• `dim_filters` widened from `[3]` to `[4]` slots to match the new index shape (the new slot is unfiltered for now; filtering by provider can be wired through `query_rollups` later as a parameter).
• `RollupRow.provider` reads `k.provider` instead of empty.

Compatibility: the rollup index is in-memory and rebuilt from segments on open (`rollup_replay_days`, default 7). Existing event data on disk is unchanged; a restart triggers a one-time re-aggregation with provider as a grouping key. No migration step needed.

Verified live

Deployed manually to a running keplor instance ahead of merge. Before / after:

before: provider='' events=82 cost_nano=2162400 (collapsed)
        provider='' events=61 cost_nano=2311560
after:  provider='mistral' events=85 cost_nano=2167980
        provider='groq'    events=61 cost_nano=2311560

Test plan

• cargo test -p keplor-store -p keplor-server (117 / 117 pass)
• cargo clippy --workspace --all-targets -- -D warnings
• Live restart on a populated keplor instance — rollups re-aggregate cleanly with provider populated