Per-table count limits in a binary decoder don't bound its memory: charge each entry's in-memory size before allocating

Pattern seen in a C++ decoder for a compressed, varint-coded table format. Every count was checked against its own limit and against the bytes left (count <= bytesLeft / minEncodedBytes) before resize(), and the decompressed size was checked against a decode budget. A fuzz/probe review still found that a ~240 B zstd frame decompressing to ~130-330 KB of tables (well under a 1 MiB budget) produced 5-7 MB of live heap. The reason: each entry is 1-5 encoded bytes but 50-100+ bytes in memory (structs holding several std::vector members, each op with its own operand vector). So the bytes-left check bounds the entry count, not the memory.

Fix that worked: give the reader a budget and a used counter, and make the count helper take an elemBytes argument (sizeof the in-memory element, plus any per-entry scratch the later validation pass allocates), charging n * elemBytes before resize. Also charge variable parts (inner vectors, payload planes, derived state the consumer will allocate from the entry's fields) at the point they're read, and reserve() exact sizes instead of push_back growth so the charge equals the allocation. Pass budget - decompressedSize to the table decoder. After this, the same inputs were refused for the budget with peak heap 0.13-0.77 MB under a 1 MiB budget.

Test that catches it: install a counting operator new (track live bytes and peak, with no hard cap so the red run doesn't throw bad_alloc), build tiny compressed inputs that expand into many cheap-to-encode entries (many empty records, one record with 2^17 one-byte ops, many small sub-records), and assert both that decode is refused and that peak heap above baseline <= budget * 1.25. Also assert that the same input decodes under a large budget, so the test isn't passing on an unrelated refusal.

Related lesson from the same review: when the bug is an out-of-bounds read, an ASan-only red test isn't red in the plain build. Also assert the refusal *reason* (the error message names the right check), so the test fails in both builds.