examples: drive both examples with the bandit module

[sshaplygin]

Both examples carried their own ~120-line Thompson Sampling adapter over
stitchfix/mab, written before this repository shipped a bandit of its own.
The first thing a reader met was a page of glue rather than the cache, and it
pulled gonum and golang.org/x/exp into example modules that need neither.
Both now call bandit.NewThompson.

The basic example's adapter was broken

Its GetRewards ranged the per-arm map and used the PolicyType as a slice
index:

for i, arm := range crs.stats {        // i is a PolicyType, not an index
    distributions[i] = mab.Beta(...)   // LFU == 2, len(distributions) == 2

so the first SelectPolicy after an epoch elapsed panicked with an index out
of range. It compiled, and was never run.

Both examples take -smoke

Build the cache, drive 100 keys through it, report, exit 0, without serving.
CI runs it for both examples, and for migration across all three strategies —
which demonstrates the strategies actually differ rather than only that the
binary starts:

strategy=cold     LRU -> LFU    0/100 keys readable after the switch
strategy=warm     LRU -> LFU  100/100 keys readable after the switch
strategy=gradual  LRU -> LFU  100/100 keys readable after the switch

Neither example's policy set, epoch settings nor cache construction changed;
only the bandit did.

Docs

docs/design.md loses the pointer to the removed adapter and gains a short
illustrative skeleton for wrapping an outside bandit — checked line by line
against interfaces.go — with the two traps that catch people: consuming
randomness while ranging a map breaks a seeded run, and an arm that saw no
requests is not an arm that scored zero. docs/getting-started.md drops
stitchfix/mab from its references.

Verification

make all                                  -> Release checks passed
git ls-files | xargs grep -l stitchfix    -> no matches
grep -c 'gonum\|x/exp' examples/*/go.sum  -> 0
go run . -smoke                           -> exit 0, both examples