Use linear time algorithm to inject stack height metering

[athei]

The current algorithm to inject the preamble and postamble for call instructions has quadratic complexity:

O(number_call_instructions_in_function * number_instructions_in_function)

This is because the splice function needs to copy around the Vec of instructions for every call instruction. A function can consist of only calls. This was evident when experimenting with different benchmarks. This PR replaces this by a linear algorithm. The benchmark in question went from 20secs to 1sec runtime.

Please note that leaving this unfixed is a DoS vector when the stack height metering is applied to untrusted code.

[pepyakin]

This looks sensible.