Impl Extend for LiteMap and avoid quadratic behavior in from_iter and deserialize

[arthurprs]

Adds lm_extend to the StoreMut trait. The Vec implementation is optimized to have O(N) performance on sorted inputs and an asymptotic worst case of O(NLog(N)), effectively avoiding quadratic worst cases.

This PR takes advantage of the new extend optimization and changes the FromIterator and Deserialize implementations to also avoid quadratic worst-case.

Before

litemap/from_iter_rand/small
                        time:   [916.60 ns 941.77 ns 965.82 ns]
litemap/from_iter_rand/large
                        time:   [2.3779 s 2.3863 s 2.3946 s]
litemap/from_iter_sorted/small
                        time:   [84.010 ns 84.084 ns 84.155 ns]
litemap/from_iter_sorted/large
                        time:   [725.31 µs 739.88 µs 755.20 µs]

After

litemap/from_iter_rand/small
                        time:   [814.95 ns 830.50 ns 845.94 ns]
litemap/from_iter_rand/large
                        time:   [42.048 ms 43.138 ms 44.337 ms]
litemap/from_iter_sorted/small
                        time:   [76.707 ns 76.759 ns 76.808 ns]
litemap/from_iter_sorted/large
                        time:   [640.42 µs 652.98 µs 668.38 µs]
litemap/extend_rand/large
                        time:   [108.48 ms 110.88 ms 113.52 ms]
litemap/extend_rand_dups/large
                        time:   [173.80 ms 175.58 ms 177.54 ms]
litemap/extend_from_litemap_rand/large
                        time:   [2.0401 s 2.0488 s 2.0575 s]

This is a followup of #6068

[arthurprs]

I considered making this based on StoreMut::lm_push and Store::lm_into_iter by adding the StoreIntoIterator bound to R, but that would be a breaking change.

[arthurprs]

Note for reviewers: The addition of alloc was theoretically a breaking change for people using "no default features" so we may want to bump litemap to 0.8.