tracing(csharp): introduce Connection/Statement lifetime activities for hierarchical traces

[eric-wang-1990]

Motivation

Issue #477 shows every top-level driver call creates a fresh root Activity
with a new TraceId. The MemoryStress baseline measured 12,316 of 12,958
TraceIds (95%) as single-span throwaway ReadNextRecordBatchAsync roots,
breaking trace-tree visualization on fleet dashboards. The
ConcurrentDisposeTests evidence confirms every
DatabricksCompositeReader.Dispose and HiveServer2Connection.DisposeClient
span has ParentSpanId=0, so even within one connection Dispose cannot be
linked back to the Execute that spawned the cleanup.

Design

Mirror the existing object hierarchy (Connection owns Statement owns calls)
in the Activity tree by introducing two long-lived "lifetime" Activities:

DatabricksConnection.Lifetime (root)
├─ HiveServer2Connection.OpenAsync
├─ ApplyServerSidePropertiesAsync
├─ DatabricksStatement.Lifetime
│   ├─ HiveServer2Statement.ExecuteStatementAsync
│   ├─ ReadNextRecordBatchAsync × N
│   └─ DatabricksCompositeReader.Dispose
├─ DatabricksStatement.Lifetime  (next statement, same connection)
└─ HiveServer2Connection.DisposeClient

Existing per-call activities need no code change: StartActivity already
uses Activity.Current as the parent, and the lifetime Activities are on
the stack throughout the synchronous Database.Connect flow and all
subsequent operations.

Implementation choices

Static ActivitySource for the lifetime spans. The per-instance
ActivitySource held by ActivityTrace is disposed inside
TracingConnection.Dispose before DatabricksConnection.Dispose returns;
listeners stop receiving notifications once the source is disposed. A
process-wide static readonly source survives connection disposal and keeps
notifying listeners. (This was caught by the failing test — first attempt
used the per-instance source and the ConnectionLifetime Stop() silently
dropped.)

Explicit parent context for Statement → Connection.
Activity.Current is AsyncLocal-scoped. If CreateStatement runs on a
different async continuation than where the connection was opened, naïve
reliance on Activity.Current would not see the connection's Lifetime.
DatabricksStatement therefore passes connection.LifetimeActivity?.Context
explicitly to StartActivity, making parenting robust regardless of async
context.

Stop order matters. ConnectionLifetime is stopped after base.Dispose
runs DisposeClient + emits DELETE_SESSION telemetry, so those tail spans
chain underneath the lifetime root. Statement lifetime is stopped after
base.Dispose runs the CLOSE_STATEMENT telemetry. Both Stop() calls are
idempotent via a _lifetimeActivityStopped guard.

Lifetime opens after ValidateProperties. If the parser throws (e.g.
ArgumentException from a bad adbc.databricks.fetch_heartbeat_interval
value), no half-open Activity is leaked — the constructor never reaches the
StartActivity call.

Tests

• ConnectionStatementLifetimeTraceTests.ConnectionLifetime_ParentsStatementAndCalls_Issue477
  drives a SELECT 1 workflow through a Thrift connection, captures every
  Activity emitted on AdbcDrivers.*, and asserts:

  1. DatabricksConnection.Lifetime exists and is a root span
  2. DatabricksStatement.Lifetime is a child of ConnectionLifetime,
     same TraceId
  3. At least one per-call activity is a child of ConnectionLifetime
     (proves the lifetime is on the Activity.Current stack during sync
     Database.Connect)
  4. At least one per-call activity is a child of StatementLifetime
     (proves the explicit parent context works)
  5. A single TraceId covers every activity on the
     AdbcDrivers.Databricks source

  Before the fix all 5 assertions fail (no Lifetime spans exist at all,
  and the captured 24 activities split across 9 distinct TraceIds — the
  exact tracing(csharp): no driver-session root span — every top-level call creates a fresh TraceId #477 baseline).

  After the fix every span shares one TraceId and the parent relationships
  match the diagram above.

Regression sweep

All targeted regressions pass against a live pecotesting Thrift connection:

Filter	Result
DatabricksConnectionTest	93 passed (0 failed)
StatementTests	94 passed (3 skipped)
CloseOperationE2ETest	3 passed
TelemetryBaselineTests	10 passed
ConcurrentDisposeTests	3 passed
Telemetry (broader sweep)	414 passed (7 skip)

No existing test required modification — the lifetime spans don't disturb
any assertion that depended on per-call activity names, status codes,
events, or tags. Tests that checked "is this span a root" continue to
hold because they only checked specific operation names; the new Lifetime
spans are net-additive.

Files touched

• csharp/src/DatabricksConnection.cs (+45 net)
• csharp/src/DatabricksStatement.cs (+27 net)
• csharp/test/E2E/Telemetry/ConnectionStatementLifetimeTraceTests.cs (+255, new file)

No submodule changes. No new public API surface beyond internal LifetimeActivity
on the connection for the statement constructor to consume.

Closes #477

This pull request and its description were written by Isaac.