No need to forward in Ptr::at(i)

Forwarding hurt performance a lot in bench-at, but I'm not sure why that is --
if anything it should be the opposite, since start(RA_FW(i)) is a nop when i is
a temp.

Author: lloda

ra/expr.hh

@@ -361,7 +361,7 @@ struct Ptr final
     constexpr static bool keep(dim_t st, int z, int j) requires (is_constant<S>) { return st*step(z)==step(j); }
     constexpr bool keep(dim_t st, int z, int j) const requires (!is_constant<S>) { return st*step(z)==step(j); }
     constexpr void adv(rank_t k, dim_t d) { mov(step(k)*d); }
-    constexpr decltype(*cp) at(auto && i) const { return *indexer(*this, cp, start(RA_FW(i))); } // iter's not view's
+    constexpr decltype(*cp) at(auto const & i) const { return *indexer(*this, cp, start(i)); } // iter's not view's
     constexpr decltype(*cp) operator*() const { return *cp; }
     constexpr auto save() const { return cp; }
     constexpr void load(P p) { cp = p; }

test/genfrom.cc

@@ -16,7 +16,11 @@ RA_CK(inside(i, len(k)), "Bad index iota [", i.n, " ", i.cp.i, " ", i.s, "] in l
 
 TODO
 * [ ] optimize() applies to ops with general iota
-* [ ] Ptr is a slice (but being Slice and Iterator at the same time is trouble)
+* [ ] Ptr is a slice (but being slice and iterator at the same time is trouble).
+  - [ ] There is conceptual confusion about view vs iterator overall. For example, view.at() and
+    iter.at() behave differently (and ra::at(a, i) uses a.at). Maybe these should have different
+    names, because he flexibility in view.at() can be expensive. Also, from(a ...) requires view a,
+    but iotas are (currently) only iterators.
 * [ ] review '1-past is ok but 1-before is not' change in fromb.cc
 * [ ] index checks
   - [X] for scalar