Improve radis import speed

[erwanp]

Description

Improved RADIS import time by delaying the import of packages not necessarily used immediatly (or ever).
(also added a test for Cantera calculations on the default Travis test environment)

Tested with tuna

python -X importtime -c "import radis" 2> radimport.log
tuna radimport.log

Before :

After:

So import is 75% faster (7s -> 4s) !

Delayed :

• Cantera (rarely used)
• Matplolib (maybe not necessarily used in the future ; with a Bokeh/Plotly backend Add plotly/bokeh backend for plots #179 )
• Astroquery (used only if fetching from HITRAN)

What's left (not for this PR)

• Astropy units . ; it's hard to delay Astropy because we need Astropy.units to know if user give Astropy.units. A way would be to "duck-test" it; i.e. evaluate a property.
• Numba & underlying llvm : is used here and there; in particular for Whiting's JIT. Would be a lot of work to delay this ; and may leave some bugs undected.
• Pandas / Scipy / Numpy are required anyway.

Notice the time it takes Numba to compute Whiting's JIT.

[codecov-commenter]

Codecov Report

Merging #309 (a654d6b) into develop (a85f2fe) will decrease coverage by 0.20%.
The diff coverage is 14.70%.

@@             Coverage Diff             @@
##           develop     #309      +/-   ##
===========================================
- Coverage    74.51%   74.31%   -0.21%     
===========================================
  Files          121      122       +1     
  Lines        14806    14823      +17     
===========================================
- Hits         11032    11015      -17     
- Misses        3774     3808      +34     

[encrypted-soul]

@erwanp Does the tool render different maps on different machines? I fetched this branch and ran tuna and I seemed to receive a different map? But there was significant improvement with this branch though.

[encrypted-soul]

Basically import packages when necessary (?)

[erwanp]

Yes ! Imports within functions is harder to maintain so you don't want to do that for things you'll have to load anyway.

For astropy.units it didn't really work; as it's still imported at one point to check the inputs.