mcp: read a colon-less SSE line as a field with an empty value

[gu-feng418]

The SSE parsing rules read a line with no : as a field name whose value is
the empty string, so an unknown such line is ignored and a bare data line
contributes an empty data line. scanEventsLimited instead reported
errMalformedEvent for it, and processStream treats that error as terminal
and calls c.fail, so one stray line ended the whole client session. The
legacy SSEClientTransport shares the same scanner.

This matters for gateways and proxies that interleave a non-standard keepalive
line into an event stream: the connection is unusable after a single line that
the spec defines as valid.

The scanner now takes the whole line as the field name and an empty value, and
the existing switch below keeps ignoring unrecognized names. Comment lines
(starting with :) and the size cap are untouched, so errMalformedEvent is
still reported for an event that exceeds maxEventSize.

Two cases in mcp/event_test.go asserted the old rejection, so they now assert
the spec reading; TestStreamableClientConnect_ColonlessSSELine covers the
session-level repro from the issue.

Fixes #1373

[chrikrah]

@gu-feng418 this is right, and I would merge it. The session-level repro from #1373 fails on the merge base and passes at f2ce77c, with the full suite green on both trees.

# go 1.27.1 linux/amd64; head f2ce77c, merge base 944ac19
$ go test ./mcp/ -run TestProbeColonlessSSELine -count=1 -v   # base 944ac19
PROBE extra="heartbeat"
PROBE extra="data"
--- FAIL: TestProbeColonlessSSELine/heartbeat
--- FAIL: TestProbeColonlessSSELine/data
FAIL	github.com/modelcontextprotocol/go-sdk/mcp

$ go test ./mcp/ -run 'TestProbeColonlessSSELine|TestStreamableClientConnect_ColonlessSSELine' -count=1 -v   # head f2ce77c
--- PASS: TestProbeColonlessSSELine
--- PASS: TestStreamableClientConnect_ColonlessSSELine
ok  	github.com/modelcontextprotocol/go-sdk/mcp	0.020s

$ go build ./... && go test ./... -count=1
# head f2ce77c: exit 0, 14 packages ok
# base 944ac19: exit 0, 14 packages ok

non-blocking: SSEClientTransport shares this scanner, and the body says so, but the new session test covers only StreamableClientTransport. One case on the legacy transport would pin that half of the claim.

@guglielmo-san #1373 asked whether the scanner should follow the SSE rules, and the two cases you rewrote in mcp/event_test.go are that decision. Do you want the spec reading, or should a colon-less line keep failing?