[C++] Bound aligned batch decodes by remaining TsBlock capacity (#1045)

* [C++] Bound aligned batch decodes by remaining TsBlock capacity

An aligned table query could hang forever in the first ResultSet::next()
when the internal TsBlock could not hold the decoder's fixed 129-row batch.
decode_tv_batch() required a full batch before consuming any input, so a
small block never filled, E_OVERFLOW was reported as an empty success, and
the scan iterator retried the same decode without progress.

Cap each pass at min(BATCH, remaining) as the multi-column path and
ChunkReader already do, and stop reporting an empty E_OVERFLOW as success:
a block that cannot hold a single row now surfaces E_INVALID_ARG instead of
spinning silently.

Closes #1035

* [C++] Free zero-sized ByteBuffer allocations

ByteBuffer::init(0) still goes through mem_alloc, which allocates an
8-byte header block, but the destructor only released the buffer when
reserved_size_ was non-zero. A TsBlock built with a capacity below one
row therefore leaked one block per column, which LeakSanitizer reported
as a failure of the new small-capacity regression test on Ubuntu.

Free whenever the pointer is set; data_ is only ever assigned by
mem_alloc/mem_realloc, so this cannot release foreign memory.

* [C++] Free zero-sized BitMap allocations

Same defect as the ByteBuffer fix: BitMap::init(0) rounds its bitmap size
to zero but still allocates an 8-byte header block, and the destructor
only released it when size_ was non-zero. A TsBlock built with a capacity
below one row leaked the null bitmap of every column, which LeakSanitizer
reported on the new small-capacity regression test.

Free on the pointer instead of the size, matching ByteBuffer.