Add ReasoningEffort.Max

[AkiKurisu]

Adds a Max level to ReasoningEffort, one tier above ExtraHigh. Both providers with
first-party IChatClient adapters now expose it: OpenAI added "max" to
ResponseReasoningEffortLevel in
openai-dotnet#1289, and Anthropic
12.42.0 already exposes Max on Anthropic.Models.Messages.Effort. Since ReasoningOptions
is sealed with only Effort and Output, there is currently no provider-neutral way to
request the top tier.

Implements the API proposed in #7715.

Changes

• ReasoningEffort.cs: Add Max. Reword the ExtraHigh summary, which claimed
  "Maximum reasoning for the most demanding tasks".
• OpenAIChatClient.cs / OpenAIResponsesChatClient.cs: Map ReasoningEffort.Max
  to new ChatReasoningEffortLevel("max") / new ResponseReasoningEffortLevel("max"),
  using the string constructor as "xhigh" does today (Map ReasoningEffort.None and ExtraHigh to "none" and "xhigh" in OpenAI IChatClient implementations #7319). These can move to the
  generated .Max property once a release containing openai-dotnet#1289 is picked up.
• Microsoft.Extensions.AI.Abstractions.json: Add the new stable field to the API
  baseline.
• Tests: Add Max as [InlineData] on the existing parameterized theories in
  ReasoningOptionsTests, OpenAIChatClientTests, and OpenAIResponseClientTests. The
  latter two assert the emitted "max" wire value.

Microsoft Reviewers: Open in CodeFlow

[AkiKurisu]

@dotnet-policy-service agree

[AkiKurisu]

@dotnet-policy-service agree