Validate STACKER init parameters against overflow and tensor sizes

[adilburaksen]

BUG=out-of-bounds access in STACKER kernel from unvalidated init parameters

What

StackerInit reads num_channels, stacker_left_context, stacker_right_context
and stacker_step from the init flexbuffer and computes:

buffer_size = num_channels * (stacker_left_context + stacker_right_context + 1);
step_size   = num_channels * stacker_step;

StackerPrepare validated only tensor ranks/types — never these parameters. A
crafted model (e.g. stacker_right_context = -1, making buffer_size = 0, or
values whose product overflows) produces a zero/undersized circular buffer.
StackerEval then calls CircularBufferWrite / CircularBufferExtend, copying
num_channels values per frame and reading buffer_size values into the output,
out of bounds.

In release builds the ASSERT(...) in CircularBufferWrite (#define ASSERT assert) is compiled out under -DNDEBUG, so this is an out-of-bounds
read/write. Confirmed with AddressSanitizer (built with -DNDEBUG):

ERROR: AddressSanitizer: global-buffer-overflow in __asan_memcpy
  #1 CircularBufferWrite   signal/src/circular_buffer.cc:97
  #2 CircularBufferExtend  signal/src/circular_buffer.cc:178
  #3 StackerEval           signal/micro/kernels/stacker.cc:139

Change

• stacker.cc: StackerPrepare validates num_channels > 0, contexts >= 0,
  stacker_step > 0, num_channels <= input size, that the buffer_size product
  does not overflow, and that the output is large enough.
• stacker_test.cc: regression test that the malformed parameters are rejected.

No change for valid models.