generate lax.slice instead of lax.gather for x[<int>] or x[:<int>]

[mattjj]

Before, on CPU:

After, on CPU:

Following up on #11866.

I don't know if these small compile time differences matter, but hey, roofshots! And the jaxpr pretty-printing win is enough for me.