⚡️ Avoid creating temporary expression in matrix multiplications

[blegat]

With the benchmark in #1403 (comment):
Before:

julia> @time x*A;
0.559247 seconds (8.21 M allocations: 668.261 MiB, 40.98% gc time)

julia> @time A*x*A;
33.784054 seconds (14.60 M allocations: 9.227 GiB, 51.83% gc time)

After:

julia> @time x*A;
0.153061 seconds (210.03 k allocations: 73.168 MiB)

julia> @time A*x*A;
16.037227 seconds (630.06 k allocations: 6.126 GiB, 14.55% gc time)

[mlubin]

Could we add this case to test/perf?