Julia: performance gotcha in scalar `.*=`

[JeffBezanson]

The bench_eu_devec benchmark got about 20x slower due to this line:

            S[j] .*= exp(t1 + t2*randn())

Changing .*= to *= fixes it. There is a lot of overhead here from making a view and doing a couple function calls (compared to an inlined multiply and store). Here are some ideas:

1. Make it an error to make a single-element dotview. Might be better to get an error telling you to remove the dot instead of mysterious bad performance.
2. Lower a[i] .op= x to dotupdate(a, i, op, x), and write a method for i::Int so this case can be fully inlined.

[vchuravy]

This seems to be a consequence of #19799?

[stevengj]

@vchuravy, the issue here is that a[i] .= ... creates view(a, i), which is a 0-dimensional subarray for scalars, which predates that PR.

+1 for dotupdate or similar. We've generally tried to make dot operations efficient for scalar operands, to facilitate writing generic code.