Small fixups to #600

[martinholters]

• Fix multi-column _small_filt_fir! for ndims(x)>2.
• Consistently require mutli-column state to match input layout. (A single state as input is still supported.) For a 2D input -- the only shape properly supported prior to Fix multi-column filt for ndims(x)>2 #600 -- this doesn't change anything.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.90%. Comparing base (1be88b5) to head (e6ffce8).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #601      +/-   ##
==========================================
- Coverage   97.90%   97.90%   -0.01%     
==========================================
  Files          19       19              
  Lines        3249     3246       -3     
==========================================
- Hits         3181     3178       -3     
  Misses         68       68              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wheeheee]

I think the size(si, 2) != ncols check should also be changed.

[martinholters]

Indeed, I think the case of passing in si with per-column-state doesn't work even with these changes as it is unclear whether the state for a, say, $L \times M_1 \times M_2$ input should be $N \times (M_1 \cdot M_2)$ or $N \times M_1 \times M_2$, and it is assumed differently in different places. I'll add tests that cover these and extend the fixup to always use the latter.

[martinholters]

The checks base on trailingsize were more permissive, but the indexing below would then fail if the sizes didn't match, so test up here. (Also, trailingsize was accessing a Base internal.)