Avoid recursion in _banded_muladd! for vectors

[jishnub]

We may avoid recursion in _banded_muladd! and directly call _banded_gbmv!instead. This makes type-inference easier, and reduces TTFX.
On master

julia> B = BandedMatrix(0=>ones(1000));

julia> v = zeros(size(B,2));

julia> @time B * v;
  1.007807 seconds (2.95 M allocations: 193.016 MiB, 7.89% gc time, 182.64% compilation time)

This PR

julia> @time B * v;
  0.294613 seconds (647.92 k allocations: 42.859 MiB, 5.88% gc time, 151.79% compilation time)

Tests for BlockBandedMatrices, ApproxFunBase, ApproxFunOrthogonalPolynomials and ApproxFun pass locally for me with this PR. This also resolves the issue with precompiling B * v in #283 (comment)

[codecov]

Codecov Report

Base: 80.47% // Head: 80.55% // Increases project coverage by +0.07% 🎉

Coverage data is based on head (6a85323) compared to base (7315691).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #293      +/-   ##
==========================================
+ Coverage   80.47%   80.55%   +0.07%     
==========================================
  Files          22       23       +1     
  Lines        3262     3270       +8     
==========================================
+ Hits         2625     2634       +9     
+ Misses        637      636       -1     

Impacted Files	Coverage Δ
src/generic/matmul.jl	72.93% <100.00%> (+1.39%)	⬆️
src/BandedMatrices.jl	100.00% <0.00%> (ø)
src/generic/Band.jl	71.60% <0.00%> (+0.35%)	⬆️
src/lapack.jl	86.36% <0.00%> (+0.48%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[jishnub]

Gentle bump @dlfivefifty