Reduce no-op copy writes beyond the tensor extent

Behavior

At cann/atvc source commit c1c9b60146aea55d71f039042dbe8d09c07b7428,
CalcReduceTiling accepts shape {1} with an empty reduction-axis list and
selects the no-reduction REDUCE_POLICY0. For a one-element float input, the
identity-copy path should write only the one valid output element. Instead, it
overwrites the next seven floats past the declared output extent with following
input values.

Cause

ReduceOpTemplate::CopyInput2Output() rounds the element count up to a 32-byte
boundary before copying. Its final tile then uses element-count DataCopy
operations, so a one-float tensor is copied as eight floats. This affects valid
small no-reduction tensors; an 8-float aligned tensor is copied correctly.

Fix

Keep the logical element count exact and use byte-length DataCopyPad for the
final input and output transfers. Add a host tiling assertion for the empty-axis
no-reduction case and simulator regressions for the one-float extent and the
8-float aligned control.

Scope

The baseline failure and fixed regression were reproduced in the 910B
TikicpuLib CPU simulator. No physical NPU execution was performed.

Validation

At source commit c1c9b60146aea55d71f039042dbe8d09c07b7428, the 910B TikicpuLib
simulator reproduced an out-of-extent write for a one-float no-reduction copy:
the seven output guard values after the valid element were overwritten. The
8-element aligned control passed.

After the fix, the one-element and aligned copy regressions passed (2/2), the
host tiling regression selected REDUCE_POLICY0 with one output element, and
the full ATVC unit suite passed (13/13). The patch keeps the logical element
count exact and uses byte-length DataCopyPad for the final tile.

No physical NPU execution was performed. The simulator coverage is limited to
the tested float copy cases; other data types, platforms, and performance were
not evaluated.

关联 PR: https://gitcode.com/cann/atvc/merge_requests/53