simplify and slightly improve memorynew inference

[oscardssmith]

while investigating some missed optimizations in #56847, @gbaraldi and I realized that copy(::Array) was using jl_genericmemory_copy_slice rather than the memmove/jl_genericmemory_copyto that copyto! lowers to. This version lets us use the faster LLVM based Memory initialization, and the memove can theoretically be further optimized by LLVM (e.g. not copying elements that get over-written without ever being read).

julia> @btime copy($[1,2,3])
  15.521 ns (2 allocations: 80 bytes) # before
  12.116 ns (2 allocations: 80 bytes) #after

julia> m = Memory{Int}(undef, 3);
julia> m.=[1,2,3];
julia> @btime copy($m)
  11.013 ns (1 allocation: 48 bytes) #before
  9.042 ns (1 allocation: 48 bytes)   #after

We also optimize the memorynew type inference to make it so that getting the length of a memory with known length will propagate that length information (which is important for cases like similar/copy etc).

[oscardssmith]

this appears to be ready to merge. will do so in a few days sans objections