Another liveness hack

[lthls]

This tries to avoid exponential blow-up in liveness.
The initial idea from @chambart was to notice if we're recomputing the live set of an expression with the same arguments, in which case we can reuse the previous result.
However, there was a bug because the previous result was read from the live field of the next instruction, which can contain an initial empty set or the result of a previous run, but doesn't mean it is actually the result of the analysis.
In addition, the env parameter needs to be compared too, and it is not stored anywhere.
So I introduced a cache that remembers, for each handler, the last arguments (env and at_join) and result (before_handler) of the last time we went through it.
Of course, this doesn't help much, because the environment will likely be different each time (otherwise we wouldn't get called again), so I cached the restriction of the environment to the free continuations of the handler, which should be safe and still avoid the exponential blow-up.

I haven't run the libbigarray test yet, but will try it and report if it compiles in a reasonable amount of time.

[lthls]

Preliminary results seem to indicate that liveness works fine, but something else allocates a lot and I couldn't get past the OOM killer. I will look further.

[lthls]

So apparently, Spill has the same problem. Guess I'll go for the same solution...

[lthls]

Merging for now, as it blocks further testing, but this should still be reviewed