deep: eliminate ctxt prjn and just build-in 1-to-1

[rcoreilly]

this definitely works best and asserting it directly makes code simpler and more obvious, and also eliminates the special learning rule for context inputs which is now moot.

[rcoreilly]

this is not possible because lateral connections within the context layer itself are critical.

Also, this is one place where Act level "display only" vars are used in learning -- I just updated ActPrv = AvgM instead of ActP and that significantly improved deep_fsa learning -- AvgM contains much more information about integrated state from last trial..