Use Toeplitz-dot-Hankel for very large dimensional transforms

Toeplitz-dot-Hankel has better complexity plans so should be the default when arrays are large: Eg for cheb2leg even n = 1000 sees a signifcant speedup:

julia> n = 100; x = randn(n); @time cheb2leg(x); @time th_cheb2leg(x);
  0.000267 seconds (2 allocations: 944 bytes)
  0.000654 seconds (98 allocations: 288.531 KiB)

julia> n = 1000; x = randn(n); @time cheb2leg(x); @time th_cheb2leg(x);
  0.028686 seconds (2 allocations: 7.984 KiB)
  0.006464 seconds (99 allocations: 10.559 MiB)

julia> n = 1000; x = randn(n); @time cheb2leg(x); @time th_cheb2leg(x);
  0.028856 seconds (2 allocations: 7.984 KiB)
  0.011597 seconds (99 allocations: 10.559 MiB)

julia> n = 10_000; x = randn(n); @time cheb2leg(x); @time th_cheb2leg(x);
  0.778423 seconds (3 allocations: 78.219 KiB)
  0.103821 seconds (108 allocations: 799.524 MiB)

This was prematurely changed but reverted as there were some regressions. But also the number of allocations in th_* is exorbitant, probably because it dates back to a port of Matlab code.

For matrices I'm not seeing much improvement in the th_*, even for a 40k x 10 transform, which is suspicious....but doing a profile all the time is in the FFTs

Start by locating the cheb2leg and th_cheb2leg transform implementations and compare their behavior for the reported vector sizes. Benchmark both paths at n=100, 1,000, and 10,000, then profile the matrix case and FFTs mentioned in the issue. Done means choosing the faster path for very large arrays without reintroducing the reported regressions or excessive allocations.