v2.0.0

A major release with some breaking changes: the library no longer panics on
input it can check, and the types say more about the values they carry. Nothing
about the analysis itself changed, so the numbers you get out stay the same.

Documentation improved a lot.

Migrating

• The two lookups return an Option now, so add a ? or an unwrap():
  spectrum.freq_val_exact(1000.0) and freq_val_closest(1000.0) give
  None for a frequency outside the spectrum instead of panicking.
• Frequency and FrequencyValue are different types now. Both compare and
  calculate with f32 directly, so assert!(value > 0.85) works and most
  val() calls can go.
• Build a frequency limit with FrequencyLimit::{min,max,range} instead of
  the variants, which hold a NonNegF32 now.
• Gone for good: scaling::combined (use a closure),
  FrequencySpectrum::{median,to_map,to_mel_map,mel_val} (use data() or
  to_vec()), FrequencyLimit::{min,max} as getters (use maybe_min() and
  maybe_max()), FrequencySpectrum::new and its Default implementation
  (a spectrum comes from samples_fft_to_spectrum).
• Errors moved around: TooFewSamples and SamplesLengthNotAPowerOfTwo
  became InvalidLengthOfSamples, and InvalidSamplingRate is new.

Fewer panics

• BREAKING FrequencySpectrum::{freq_val_exact,freq_val_closest} return
  None for a frequency outside the spectrum, and for NaN, which slipped
  through the bounds check before
• BREAKING more than 32768 samples is an error instead of a panic
• BREAKING a sampling rate of zero is rejected with the new
  SpectrumAnalyzerError::InvalidSamplingRate, instead of producing a
  spectrum in which every frequency is 0 Hz
• BREAKING removed FrequencyLimit::{min,max}, which panicked for
  variants without that bound
• BREAKING FrequencySpectrum::new is internal and the Default
  implementation is gone; the empty spectrum it produced made every getter
  panic

Frequencies and values have their own types

• BREAKING a Frequency is a NonNegF32 and a FrequencyValue a
  FiniteF32, so the two cannot be mixed up. The difference is real: a
  frequency is never negative, a value turns negative as soon as
  scale_20_times_log10 touches it
• BREAKING FrequencyLimit::{Min,Max,Range} hold a NonNegF32, which
  makes a negative or non-regular limit impossible to build.
  FrequencyLimitError::{NotARegularNumber,ValueBelowMinimum} are gone with
  it
• BREAKING FrequencySpectrum::frequency_resolution,
  scaling::SpectrumDataStats::{min,max,average} and
  FrequencyLimitError::{ValueAboveNyquist,InvalidRange} carry those types
  as well. SpectrumDataStats::n stays an f32, since it is there to
  divide by

A thinner spectrum

• BREAKING removed scaling::combined; it handed every function the
  statistics of the unscaled spectrum, so chaining two of them gave wrong
  results. A closure does the same job correctly
• BREAKING removed FrequencySpectrum::median and
  SpectrumDataStats::median; nothing in the library used them
• BREAKING removed FrequencySpectrum::{to_map,to_mel_map}; both used
  u32 keys, so bins that shared a key silently overwrote each other
• BREAKING removed FrequencySpectrum::mel_val; such calculations are
  easy to do on data(), and the spectrum stays thin and unopinionated

Fixed

• the Hamming and Blackman-Harris windows used the symmetric form (dividing
  by N - 1) while the Hann window used the periodic one (dividing by N).
  All of them use the periodic form now, which is the right one for FFT
  analysis. The coefficients change marginally and the coherent gains are
  exact
• scale_20_times_log10 mapped 0.0 to 0 dB, which ranked silence above
  every quieter bin. Values are clamped to -100 dB at minimum now
• the descriptions of FrequencyLimit::Min and FrequencyLimit::Max were
  swapped

Performance

• spectrum creation no longer allocates and scans a working buffer for the
  median, which makes it 20-30% faster depending on the number of samples

Documentation

• explained what the frequency values are and how they relate to the input
  signal, including the expected range of the samples
• added guidance on which window function and which scaling function to pick
• documented the coherent gain of each window, and the difference between the
  periodic and the symmetric form
• clarified divide_by_N against divide_by_N_sqrt, and the statistics a
  scaling function receives
• fixed the frequency resolution formula; it is sample_rate / N, not
  sample_rate / (N / 2)

What's Changed

• docs: clarify frequency values, scaling, and windows by @phip1611 in #85
• scaling: simplify the API and add guidance on picking settings by @phip1611 in #86
• spectrum: remove the map conversions and fix doc statements by @phip1611 in #87
• windows: use the periodic form and misc doc fixes by @phip1611 in #88
• spectrum: tighten the public API and fix the CI setup by @phip1611 in #89
• doc: streamline and remove outdated by @phip1611 in #90
• lib: replace the remaining panics with Option and Result by @phip1611 in #91
• document the expected range of samples and FiniteF32 helper by @phip1611 in #92
• lib: say what the one remaining panic means by @phip1611 in #93

Full Changelog: v1.9.0...v2.0.0