spectra as 32-bit instead of 64-bit

[sbailey]

Update io.write_frame() to write 32-bit floats, and update io.read_frame() to cast back to 64-bit double for internal calculations. This is a very simple and easily reproducible form of lossy data compression.
Related: We had a long discussion on the desi-data mailing list about the pros and cons of this vs. FPACK tile compression. FPACK can do somewhat better in terms of compression vs. precision lost, but the basic float vs. double compression seems "good enough" for now.

This is similar to what is already done for io.write_image() and io.read_image().

[sbailey]

Adding to pipeline v0.0 milestone, though this could be optional for that software release.