Replace sync cache build with recorder operator

[Andrewtangtang]

Closes: #75

Summary

This PR replaces the synchronous full-table build that the optimizer used to
run on cache miss with a PhysicalCacheRecorder operator that observes the
user's own query and builds the cache as a side-effect. The first query on
a new predicate no longer pays an extra full-table scan.

An observed_vectors watermark is added to RowGroupFilter so the cache
can distinguish "observed empty" (safe to prune) from "not yet observed"
(must scan). Recorder injection is narrow: it only runs on a full cache
miss. Partial refill and scan-skipped row groups are tracked as TODOs.

Operator Hierarchy

Cache miss — optimizer injects the recorder above LogicalGet:

LogicalFilter (user's predicate)
  └─ LogicalCacheRecorder                ← new; evaluates predicate, records
       └─ LogicalGet
            ├─ column_ids: [..., ROW_ID]
            ├─ projection_ids: [..., ROW_ID]   ← surfaced for recorder
            └─ table_filters[ROW_ID] = CacheExpressionFilter

Cache hit — no recorder, filter only:

LogicalFilter (user's predicate)
  └─ LogicalGet
       ├─ column_ids: [..., ROW_ID]
       ├─ projection_ids: [...]           ← ROW_ID read but not projected
       └─ table_filters[ROW_ID] = CacheExpressionFilter

Runtime flow for the cache-miss case:

PhysicalTableScan
  │ emits chunks (post zone-map / table filter)
  ▼
PhysicalCacheRecorder::Execute
  ├─ compute (rg_idx, vec_idx) from first row_id
  ├─ expr_executor.SelectExpression(chunk)
  └─ RecordChunkObservation -> task-local bitvector + watermark
  │ chunk.Reference(input)   (pass-through)
  ▼
PhysicalFilter / downstream ops

... at query end ...

PhysicalCacheRecorder::OperatorFinalize
  ├─ merge all task-local entries -> destination
  └─ store->Upsert(key, destination)

[peterxcli]

I think the watermark management is not quite good for cache invalidation.
how about we store another bitmask to record if the vector is observed?

struct RowGroupFilter {
	array<uint64_t, BITVECTOR_ARRAY_SIZE> matching_vectors = {};
    array<uint64_t, BITVECTOR_ARRAY_SIZE> observed = {};

[Andrewtangtang]

agree.

[peterxcli]

I think we already removed the watermark?

[peterxcli]

can you use plain words to explain the logic of this function, espically the rowid_chunk_idx and rowid_column_ids_pos

[peterxcli]

our disscusion is to keep keep the recorder anyway? and thus I think we can remove the two phase optimizer(query context)? because we aren't going to do any cache lookup thing when plan rewriting, we do all lookup and upsert things inside physical recorder operator and the condition cache filter expression when they are processing the incoming data chunk?

[peterxcli]

is this case possible? if this happened means no rows in the vector is valid, and I though all invalid vector wont flow into here?

[peterxcli]

We should think this through:

1. When the recorder sees a vector that hasn’t been observed yet, it should upsert both the observed and matching_vectors bitmasks.
2. Do we actually only need observed per row group? The recorder could upsert all cache entries for every vector it sees, then mark that row group as observed.
3. The query cache expression filter should only skip a vector if that vector’s row group is marked as observed in the cache entry.