Implemented rfft!, irfft! and brfft!

[favba]

This PR supersedes #54 .
Most of the code is the same.
It provides the same features but adds the capability of calling irfft!(c::Array{Complex},d::Integer) and rfft!(r::SubArray{<:Real} which are implicitly converted to PaddedRFFTArray. This makes it much simpler to convert code using rfft and irfft to use the in-place versions of those

In order to implement those features, we must allow PaddedRFFTArray to be initialized with a complex array, in which case the real view of the array must be reinterpreted out the complex array.
Although there is a way to make reinterpreted real-to-complex array performant (see Tim Holy's comment), it doesn't seem the same trick can be used to make reinterpreted complex-to-real arrays performant.

In this PR I'm using an internal ComplexOrRealReinterpretArray that relies on loading and storing directly from pointers to the raw data, while GC.@preserveing appropriately the underlying data.

To Do:

• Deal with ambiguity of rfft!(a::PaddedRFFTArray,region=1:ndims(a)) and rfft!(a::DenseArray{<:Complex},d::Integer) in a better way
• Document the new exported functions

[codecov]

Codecov Report

Attention: Patch coverage is 85.03937% with 19 lines in your changes missing coverage. Please review.

Project coverage is 68.14%. Comparing base (5f6f315) to head (141f60d).
Report is 56 commits behind head on master.

Files with missing lines	Patch %	Lines
src/rfft!.jl	85.03%	19 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #222      +/-   ##
==========================================
+ Coverage   62.67%   68.14%   +5.46%     
==========================================
  Files           5        6       +1     
  Lines         434      565     +131     
==========================================
+ Hits          272      385     +113     
- Misses        162      180      +18     

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

[kahaaga]

Is there any progress on this PR? This would be very useful for us to have in TimeseriesSurrogates.jl, where performance of some of our methods would be improved if the in-place inverse transform irfft! would be available.