fix(mapper): unmap fused-away scans; tolerate unmapped SEMI_MASKER targets

[adsharma]

Fixes #1068.

Root cause

unordered_map::at: key not found was a use-after-free in physical plan mapping:

1. The hash-join SIP optimizer places a SEMI_MASKER (keyed e.g. on i._ID) targeting the bound-node SCAN_NODE_TABLE on the other side of the join.
2. When mapping that side, mapExtend's fused ice-disk path absorbs the child scan into the ScanRelTable and discards its separately mapped physical operator, leaving a dangling entry in logicalOpToPhysicalOpMap.
3. mapSemiMasker then dereferenced the freed pointer (observed reallocated as a different scan at the same address), so initMask's mask-table lookup threw.

Reproduced on a v0.21.0 build (fails on both EXPLAIN and execution); current HEAD only hid it via a later join-order change.

Changes

• src/processor/map/map_extend.cpp: erase the stale child-scan mapping in fused branches that drop prevOperator.
• src/processor/map/map_semi_masker.cpp: look up SEMI targets with find() instead of at(), skip per-table masks a target doesn't hold, and drop maskers that attach nothing (result-preserving pass-through).
• test/test_files/ice_disk/ice_disk_complex_queries.test: new SemiMaskerOnFusedBoundScan regression case producing the exact crashing shape.

Verification

• Issue repro on icebug-disk: crash → 111 rows, byte-identical to the native-database result.
• A/B on HEAD build: stashed fix crashes on both the issue shape and the new test query; restored fix passes.
• ice_disk e2e suite passes (except pre-existing ice_disk_large_node_scan failure from the missing ice-disk-large-test dataset in this checkout).
• clang-format-18 clean on both edited files.