fix(ENG-RUNNER-MODELSHAPE): BlockTable::move_row leaves the vacated source row populated

Row: ENG-RUNNER-MODELSHAPE

Found by re-deriving upstream 6fbbcf2151 (vllm#49757) against this tree in
#2632, wave PORTQ-1.

The defect

BlockTable::move_row copies the source row into the target and updates the
target's block count, and then stops. It never clears the row it vacated, so the
source row keeps a full copy of the block ids it no longer owns and
num_blocks_per_row[src] keeps counting them.

src/vllm/v1/worker/gpu/block_table.cpp:105-113:

void BlockTable::move_row(int src, int tgt) {
  const int num_blocks = num_blocks_per_row[static_cast<size_t>(src)];
  ...
  num_blocks_per_row[static_cast<size_t>(tgt)] = num_blocks;
}

clear_row directly above it (:95-103) already does exactly the clearing that
is missing, so the fix is to call it.

Upstream added the same clearing at vllm/v1/worker/block_table.py:144-151
@ 6fbbcf2151 (a forward reference past the pin 5559679229), for the reason
that a dummy run then wrote Mamba state through the stale row.

What is NOT claimed

The upstream symptom is not reachable here today. This tree has no
_dummy_run, and padded decode rows are built explicitly rather than by
dereferencing whatever the block table happens to hold:
src/vllm/model_executor/models/qwen3_5.cpp:10309 writes -1 into every padded
GDN state index and slot-mapping entry, and include/vllm/v1/worker/gpu/runner.h:598
rewrites the GDN group's block-table column 0 into a compact per-sequence slot
keyed on req_id. So a stale row past num_reqs is never read.

This is a defensive fix and a mirror repair, not a live defect. It is filed
because the divergence is real and cheap to close, and because the structure that
makes it unreachable is the runner's, not the block table's: a future reader of
move_row has no way to know the invariant is upheld elsewhere.

Size

~4 lines in src/vllm/v1/worker/gpu/block_table.cpp plus one red-first unit case
under tests/vllm/v1/worker/. The live caller is
src/vllm/v1/worker/gpu/input_batch.cpp:709.

Start with src/vllm/v1/worker/gpu/block_table.cpp, comparing move_row with clear_row, then inspect the live caller at src/vllm/v1/worker/gpu/input_batch.cpp:709. Add the missing source-row clearing and a red-first unit case under tests/vllm/v1/worker/; done means the vacated row has no block IDs and its block count is zero.