Improve graph build time

[slaren]

Instead of checking the list of current nodes in ggml_visit_parents() to determine if a node has already been visited, a visited flag is added to ggml_tensor.

Unfortunately this requires resetting the visited flag to reuse the tensors in a different graph, and because in some cases graphs are built in multiple steps, it is not easy to do this automatically. Currently, this is done automatically when calling ggml_graph_plan(), but there may be some cases where it can cause current ggml code to fail silently.

[slaren]

@JohannesGaessler this may explain the performance uplift that you observed in #2230 (comment).

[slaren]

Replaced the visited flag with a very simple hash table in ggml_cgraph. The performance is almost the same and it won't break old code, but it increases the size of ggml_cgraph by 70%.

[slaren]

I think this is as good as it is going to get, ultimately it is a choice between maintaining backwards compatibility or memory usage. I am inclined to prefer the hash table version, but it increases the size of a struct that is already very large. Won't be an issue after we change graphs to allocate them on the heap/context.

Some performance numbers:

commit	graph build time ms
b47b8a9	0.963
e76d630	0.642
2f9cf97	0.534
PR (flag)	0.159
PR (hash)	0.168

Some recent changes in master reduced the graph build times as well, I have no idea why.

[ggerganov]

Nice, we should implement ggml-org/ggml#299 soon.