fix: three correctness bugs in context parsing, the memory cache and the Scaleway search

[ZeikoFr]

Three independent bugs found while profiling the live search. Each commit is
self-contained and comes with a test that fails against the current behaviour.

GetContextNames panics on a context entry without a name

valueOf(contextNode, "name") returns nil when the entry is not a mapping or
has no name key, and the result was dereferenced unconditionally. One such
entry in a kubeconfig crashes the whole search with a nil pointer dereference
instead of being skipped.

The in-memory kubeconfig cache is unsynchronised

memoryCache.GetKubeconfigForPath reads and writes its map without a lock.
Nothing calls it concurrently today, so this is latent, but any caller that
retrieves several kubeconfigs at once turns it into
fatal error: concurrent map writes. The added test fails under -race.

The upstream call is deliberately made without holding the lock: it is a remote
call taking seconds and would otherwise serialise every caller.

The Scaleway search loses clusters

• ListProjects and ListClusters were called without scw.WithAllPages(), so
  the SDK requested a single page. Every project past the server-side default
  page size, and every cluster past it within a project, was silently missing
  from the results.
• A project whose cluster listing fails, typically one the credentials cannot
  read, aborted StartSearch entirely. Every project after it was never
  reported. The error is now surfaced and the search continues.
• The "failed to create Kubernetes API instance" error wrapped an err that is
  necessarily nil in that branch, so the message ended in %!w(<nil>).

Verification

gofmt, go build, go vet, golangci-lint run (0 issues) and
go test -race -count=1 ./... are clean, and every commit passes them
individually.

🤖 Generated with Claude Code