fix: SIGSEGV due to out of bounds access after delete + checkpoint

[adsharma]

Fixes: #610

Root cause

After relationship deletes followed by a checkpoint, an in-memory
CSR index entry (csrIndex->indices[nodeOffset]) can retain a stale
INVALID_ROW_IDX (UINT64_MAX) sentinel row. Passing that through
getQuotientRemainder(row, CHUNKED_NODE_GROUP_CAPACITY) produces
an out-of-range chunk-group index, which truncates to UINT32_MAX
in getGroup(idx_t) — an out-of-bounds std::vector access guarded
only by a debug DASSERT (a no-op in release), so callers dereference
a null/OOB group.

The sentinel survives because checkpointInMemAndOnDisk calls
collectLeafRegionsAndCSRLength (which runs
collectInMemRegionChangesAndUpdateHeaderLength on all leaf regions,
writing INVALID_ROW_IDX via setInvalid(i)), but then takes an
early-return path (line 538, regionsToCheckpoint.empty()) without
calling finalizeCheckpoint — so the index is left with stale
sentinels and the chunked groups are not cleared.

Fix

src/storage/table/csr_node_group.cpp — bounds-check chunkIdx against
chunkedGroups.getNumGroups(lock) before every unguarded getGroup
call site, treating stale rows as skip/deleted (consistent with
the two already-guarded sites at lines 851 and 1060). The 6 fixed
sites:

1. scanCommittedInMemSequential → return empty result
2. scanCommittedInMemRandom → skip row
3. update (COMMITTED_IN_MEMORY) → no-op
4. delete_ (COMMITTED_IN_MEMORY) → no-op (return false)
5. collectInMemRegionChangesAndUpdateHeaderLength → treat as deleted, set invalid
6. populateCSRLengthInMemOnly → treat as deleted, decrement length