proposal: allow specifying keys that might not be implemented for CUE_EXPERIMENT

[rogpeppe]

In #3689, it's stated that failing when CUE_EXPERIMENT contains unknown keys is a problem:

As @mpvl points out, this is potentially a problem, because being able to enable or disable a number of experiments across CUE versions - even if some very old or very new versions do not support them - is useful to compare behavior, such as when investigating or bisecting bugs.

However, it seems to me that this is only a problem in this specific kind of situation.
In most other situations, it's actually good that we fail on unknown keys because
setting CUE_EXPERIMENT indicates an intention and if something is misspelled,
that's almost certainly an error.

I suggest that when we want this kind of behaviour (not failing on unknown keys), we will always know that
we want this behaviour and are in a position to do something slightly different to work around the error.

I propose that we support a special syntax that indicates that a given key should not
result in an error when it's unknown. It would still result in an error if an invalid
value is specified for a key that is recognized.

Straw-man syntax: a ? after the key:

Examples:

# Do not fail if evalv3 is not recognized but _do_ fail if embed is not.
CUE_EXPERIMENT='evalv3?,embed=1'

# Do not fail if embed is not recognized, but _do_ fail if
# embed is recognized but is not allowed to be set to false.
CUE_EXPERIMENT='embed?=0'

We might also support a mode that allows unrecognized values for known keys too, say ??:

# Set embed to false or ignore if false is not allowed.
CUE_EXPERIMENT=`embed??=1`