Validate output size in FILTER_BANK_LOG kernel

[adilburaksen]

BUG=out-of-bounds write in FILTER_BANK_LOG kernel from unchecked output size

What

FilterBankLogEval takes num_channels from the input tensor
(num_channels = input->dims->data[0]) and FilterbankLog writes num_channels
elements to the output:

for (int i = 0; i < num_channels; ++i) { output[i] = ...; }

FilterBankLogPrepare validated only the rank and type of the input and output
tensors — never that the output is at least as large as the input. A model whose
FILTER_BANK_LOG output tensor is smaller than its input therefore causes an
out-of-bounds write (e.g. input of 32 elements, output of 1), confirmed with
AddressSanitizer:

ERROR: AddressSanitizer: stack-buffer-overflow  WRITE of size 2
  #0 tflite::tflm_signal::FilterbankLog(...)  signal/src/filter_bank_log.cc
  #1 tflite::FilterBankLogEval(...)           signal/micro/kernels/filter_bank_log.cc:97
  #2 tflite::micro::KernelRunner::Invoke()

Change

• filter_bank_log.cc: FilterBankLogPrepare now requires
  ElementCount(output) >= ElementCount(input).
• filter_bank_log_test.cc: regression test that the mismatched model is rejected.

No change for valid models.