kv-cells : fix tracking of seq_pos during cache reuse

[ggerganov]

fix #14298

We tracked the sequence positions present in the KV cells using an std::set because of the assumption that a position would only be present once at a maximum. However, during cache reuse, we do the following sequence of operations:

llama.cpp/tools/server/server.cpp

Lines 3224 to 3228 in 439d562

	const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
	llama_memory_seq_rm (llama_get_memory(ctx), slot.id, head_p, head_c);
	llama_memory_seq_add(llama_get_memory(ctx), slot.id, head_c, head_c + n_match, kv_shift);

I.e. we remove a chunk of positions in [p0, p0 + match) and then move another chunk [p1, p1 + match) into its place by adding shift = p0 - p1 to all positions in that chunk. During this "movement", depending on the order that we apply the shift, a position could occur more than once for a short time, which breaks the assumption of std::set. To fix that, we replace std::set with std::map and count the occurrences of each position, erasing the elements when the count reaches 0.