[bug]: routing: BimodalEstimator fails when channel capacity >1M sat

[morehouse]

BimodalEstimator.probabilityFormula returns zero when its parameters are on the order of ~1M sat because the exponentials calculated in BimodalEstimator.primitive become tiny and round to 0.

Initially detected by @dergoegge running the FuzzProbability fuzz target. Sample failure:

Input:

go test fuzz v1
uint64(1000000000)
uint64(300000000)
uint64(400000000)
uint64(300000000)

Trace:

--- FAIL: FuzzProbability (0.18s)
    --- FAIL: FuzzProbability (0.00s)
        probability_bimodal_test.go:708:
                Error Trace:    /workdir/lnd/routing/probability_bimodal_test.go:708
                                                        /usr/local/go/src/reflect/value.go:596
                                                        /usr/local/go/src/reflect/value.go:380
                                                        /usr/local/go/src/testing/fuzz.go:335
                Error:          Received unexpected error:
                                normalization factor is zero, capacity: 1e+09, success amount: 3e+08, fail amount: 4e+08
                Test:           FuzzProbability
                Messages:       c: 1000000000 s: 300000000 f: 400000000 a: 300000000

[bitromortac]

Thanks for the report! This is because the success and fail amount drive the bimodal model into regions where the model doesn't give good predictions and therefore has problems with numerics. I tested this with #8330, where it passed, so this should be fixed by that PR in the next release.