Specify destroy() methods and behavior around context loss and error reporting (#744)

* Specify behavior around context loss and error reporting.

Based on @mingmingtasd's work in the Chromium prototype
implementation.

For #477

* Add link for 'settled'

* Add 'is lost' check to build()

* Add 'is not lost' dfn alias

* Add "can (not) build" defn, fix throwing vs. rejecting for methods

* Make build() and compute() explicitly invoke context lost steps

* reword to reduce some indenting

* restore missing algorithm class

* Specify various destroy() methods

* Lint/wording fix

* whitespace change

* another whitespace change

* Move destroy of dependencies to 'lost' steps and other feedback

* Abort steps / reject promises when context is lost

* Separate out lost/lose steps, make dispatch() use abort-when

* Incorporate more feedback, another lint rule

* Include tensor creation in MLContext/destroy() note

* Move async 'abort when'/'if aborted' into substeps

* Update index.bs

Co-authored-by: Reilly Grant <[email protected]>

* Add suggested wording by @reillyeon for freeing graph resources on context timeline

---------

Co-authored-by: Reilly Grant <[email protected]>

Author: inexorabletash