change normalization of eigenmode sources to yield unit input power

[HomerReid]

Fixes #470. Renormalize the amplitude of eigenmode sources to yield unit power flux.

The power flux depends on the frequency width of the source. I added a width parameter to fields::get_eigenmode with a default value of 0. (If width==0 the amplitude renormalization is skipped.) This requires updating calling conventions in a couple of other places:

(1) Within add_eigenmode_source, the frequency width can in principle be fetched from the existing src input parameter, so the calling convention doesn't need to change. However, the frequency width is a private data field in gaussian_src_time and continuous_src_time, and there were no getter functions. I added a virtual getter routine get_width() to the parent src_time class, which returns 0 by default and is overridden by gaussian_src_time and continuous_src_time.

(2) I updated the _get_eigenmode python routine to accept a new optional width=0.0 argument. (Again, passing width=0.0 to get_eigenmode yields behavior unchanged from before this PR.) If a user calls get_eigenmode directly from python and wants the amplitude renormalizationthe user must specify a nonzero value forwidth`.

(3) No modification is needed for get_eigenmode_coefficients. The overall normalization of the eigenmodes cancels out of this calculation anyway, so it suffices to pass width=0.0 yielding behavior equivalent to before.

[stevengj]

I think the mode normalization should not depend upon the time profile. You need to know the whole time-profile Fourier transform anyway to use this, so you might as well include the time-profile normalization there too.

In particular, the source current looks like mode(x)*p(t), where p(t) is the time profile. My suggestion is:

• Normalize mode(x) to be unit power, just like in the mode-coefficient code. This is independent of p(t), whether it is Gaussian or whatever.

• If you have a CW source, then it should result in unit power.

• If you have a Gaussian or some other source, you take this unit power and multiply by |p̂(ω)|² to get the expected spectrum (modulo the discretization error), where p̂(ω) is a Fourier transform of p(t) with an appropriate normalization. We will update the manual to provide p̂(ω) for the Gaussian source (and maybe have a Python method for this).

[stevengj]

I just merged a PR #662 that reformats everything with clang-format, which creates lots of conflicts.

To rebase this PR onto master without having to deal manually with these formatting conflicts, follow the exact procedure given in the instructions of #662.