Fix IECOREUSD_WRITE_CONFORMANT_RENDERMAN_ATTRIBUTES bugs

[johnhaddon]

I messed up in 5e2f322, because I didn't realise the full effects of changing AttributeAlgo::nameToUSD(). Despite the function name, this was used when reading attributes too, and was also used when determining the names to use for UsdShadeShader outputs. This lead to several bugs :

• Failure to read shaders written with IECOREUSD_WRITE_CONFORMANT_RENDERMAN_ATTRIBUTES=0.
• Bad output names when writing shaders with IECOREUSD_WRITE_CONFORMANT_RENDERMAN_ATTRIBUTES=1.
• Failure to read renderman attributes written with a different value of IECOREUSD_WRITE_CONFORMANT_RENDERMAN_ATTRIBUTES.

All this stuff is well overdue for a complete simplification where we only write attributes as primvars, but since we need this for 1.6.
x, I've made the most minimal change I can that I think is logical and fixes the problems. The fundamental issue is that because we want to support legacy attributes too, and nameToUSD() is used for reading, there is not a 1-to-1 mapping of Cortex names to USD names. This is reflected in the new internal usdNames() function which lists the relevant attribute names in the order we'd prefer to use them.

[murraystevenson]

Thanks John, the fixes and improved test coverage look good to me. I agree we're overdue an overhaul here...