[WIP] Use Transient HashMaps to cache wasmi modules for execution.

[gnunicorn]

refs #257

However, on our test suite (with fairly little code to actually run), the CPU drops but overall execution time increased (I suspect the hashing and mutex create some overhead, which in these really small cases take their toll):

perf stat cargo test -p substrate-executor

Performance counter stats for 'cargo test -p substrate-executor':

---- before:

      1900,661621      task-clock:u (msec)       #    1,664 CPUs utilized          
                0      context-switches:u        #    0,000 K/sec                  
                0      cpu-migrations:u          #    0,000 K/sec                  
           34.429      page-faults:u             #    0,018 M/sec                  
    5.850.197.166      cycles:u                  #    3,078 GHz                    
    8.493.281.415      instructions:u            #    1,45  insn per cycle         
    1.634.226.862      branches:u                #  859,820 M/sec                  
       10.494.673      branch-misses:u           #    0,64% of all branches        

      1,142427675 seconds time elapsed

---- after:

       1498,885534      task-clock:u (msec)       #    0,995 CPUs utilized          
                 0      context-switches:u        #    0,000 K/sec                  
                 0      cpu-migrations:u          #    0,000 K/sec                  
            33.808      page-faults:u             #    0,023 M/sec                  
     4.539.469.365      cycles:u                  #    3,029 GHz                    
     7.845.320.660      instructions:u            #    1,73  insn per cycle         
     1.515.924.592      branches:u                # 1011,368 M/sec                  
        10.394.267      branch-misses:u           #    0,69% of all branches        

       1,506522808 seconds time elapsed

Also adds TransientHashMap and lazy_static as new dependencies to the executor.

[rphmeier]

codebase style with panickers is "prove or remove": either use expect with a string describing the assumptions that show why the panic won't happen, or don't potentially panic at all.

Either way it's irrelevant because we prefer parking_lot mutexes over std

[gnunicorn]

superseeded by #276 .