Update README.md and benchmarks

[jakobkummerow]

• updated native BigInt browser support in the README.md, and a few other outdated bits
• updated benchmarks to print the tested operation, and to replace the (ancient, pre-public-release) "JSB" name with "JSBI"
• updated benchmarks for native BigInts so optimizing JS compilers can't dead-code-eliminate the operation under test
• updated the benchmark generator script to be python3 compatible
• fixed the benchmark generator's ability to run stress tests

[jakobkummerow]

I wonder how future-proof this particular way of tricking the compiler is — is there any way this might one day no longer be sufficient?

I think it's reasonably future-proof, though as always it's hard to make predictions with absolute certainty :-) The only way to be 100% bulletproof would be to use/check every single computed result, but I don't want to introduce that much overhead (since we're measuring individual operations, adding an additional comparison would be ~50% of the total workload), especially as long as it isn't necessary. For some operations, like add and sub, we could work around it by turning the test into a chain of operations, i.e. using the last result as one of the two inputs. But for bitwiseAnd and bitwiseOr, that approach would quickly lead to a rather boring all-0-bits or all-1-bits result, so it's not generally applicable.

I think pragmatically, if we notice that it breaks, we'll just fix it again. Maybe JSBI will be deprecated before that happens :-)

[jakobkummerow]

Thanks for the review!